Add audio.cpp multi-task support

Add MOSS community full-checkpoint support

Implementation details:
- Add the LAION MOSS-TTS v1.5 Voice Acting 8B community model to the engine dropdown and downloader
- Discover compatible local MOSS full checkpoints and resolve architecture from config.json
- Enable the LAION Delay checkpoint in the existing LoRA training pipeline
- Reject unsupported local MOSS layouts instead of assuming Local Transformer
- Address issue #340

.gitignore

@@ -19,13 +19,13 @@ downloads/
 !utils/downloads/
 eggs/
 .eggs/
-lib/
-!engines/rvc/impl/lib/
-!engines/rvc/training/bundled_runtime/lib/
-!engines/rvc/training/bundled_runtime/lib/**
-engines/rvc/training/bundled_runtime/lib/**/__pycache__/
-engines/rvc/training/bundled_runtime/lib/**/*.py[cod]
-lib64/
+lib/
+!engines/rvc/impl/lib/
+!engines/rvc/training/bundled_runtime/lib/
+!engines/rvc/training/bundled_runtime/lib/**
+engines/rvc/training/bundled_runtime/lib/**/__pycache__/
+engines/rvc/training/bundled_runtime/lib/**/*.py[cod]
+lib64/
 parts/
 sdist/
 var/
@@ -139,6 +139,9 @@ celerybeat.pid
 
 # Environments
 .env
+.env.local
+tests/.env
+tests/.env.local
 .venv
 env/
 venv/
@@ -146,6 +149,7 @@ ENV/
 env.bak/
 venv.bak/
 comfyui_env/
+/runtimes/
 
 # Spyder project settings
 .spyderproject
@@ -242,3 +246,10 @@ tests/output/
 
 # TTS Audio Suite settings cache (user preferences)
 .inline_tag_settings_cache.json
+
+# AI Agent skills
+.agents/
+skills-lock.json
+
+# others
+node_modules/

CHANGELOG.md

@@ -5,6 +5,437 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [5.8.1] - 2026-08-11
+
+### Added
+
+- Add MOSS-TTS community voice-acting model support
+- Add the clearly labeled LAION Voice Acting 8B community model with automatic download
+- Add compatible local full-checkpoint discovery from the MOSS model folder
+- Support experimental LoRA training with the LAION community checkpoint
+
+### Changed
+
+- Improve errors for unsupported local MOSS model layouts
+## [5.8.0] - 2026-08-11
+
+### Added
+
+- Add IndexTTS 2.5 as a new version of the existing IndexTTS engine
+- Add Chinese, English, Japanese, Spanish, and Arabic generation
+- Add explicit per-segment language switching for IndexTTS 2.5
+- Add official duration-factor and text-normalization controls
+- Keep IndexTTS 2.0 available for workflows that prefer its voice resemblance
+
+### Fixed
+
+- Fix stale audio or models when switching between IndexTTS 2.0 and 2.5
+## [5.7.0] - 2026-08-10
+
+### Added
+
+- Add integrated DramaBox LoRA model training
+- Add dataset preparation and training controls for DramaBox voice adapters
+- Add live training progress and loss reporting in the Model Training panel
+- Add DramaBox LoRA loading and adjustable adapter strength for inference
+- Add a ready-to-use DramaBox LoRA training workflow and guide
+
+### Changed
+
+- Improve shared speech-clip dataset staging for model training
+## [5.6.5] - 2026-08-03
+
+### Fixed
+
+- Fix MOSS-TTS training settings in saved workflows
+- Fix existing MOSS Dataset Prep workflows loading values into the wrong fields
+- Fix invalid validation split and preparation batch size errors after updating
+- Fix MOSS training tensor shape errors caused by shifted codec settings
+## [5.6.4] - 2026-08-03
+
+### Added
+
+- Add MOSS-TTS training dataset folder support
+- Add direct loading of matching audio and transcript files from a folder
+- Support WAV, FLAC, MP3, OGG, and M4A training clips
+- Add optional recursive scanning for datasets organized into subfolders
+- Preserve existing JSONL manifest workflows
+## [5.6.3] - 2026-08-01
+
+### Changed
+
+- Improve runtime availability checks so package startup code is not executed during installation
+
+### Fixed
+
+- Fix TTS Audio Suite installer validation failures
+- Fix ComfyUI Desktop installation failing on supported PyTorch and TorchAudio combinations
+## [5.6.2] - 2026-07-30
+
+### Changed
+
+- Improve F5-TTS fallback so the standard PyTorch attention backend continues working
+
+### Fixed
+
+- Fix F5-TTS failing to load with incomplete FlashAttention installations
+- Fix F5-TTS startup crashes when optional FlashAttention components are missing
+## [5.6.1] - 2026-07-30
+
+### Fixed
+
+- Fix Fish Audio S2 installation in headless Linux environments
+- Fix missing optional audio libraries preventing Fish Audio S2 setup
+- Improve Linux and macOS dependency warnings so core TTS installation continues
+- Correct Fedora package installation guidance
+## [5.6.0] - 2026-07-25
+
+### Added
+
+- Add DramaBox expressive TTS and ChatterBox V3 support
+- Add DramaBox scene prompting, character switching, prompt templates, and negative prompting
+- Add DramaBox native SRT duration targeting and generation-duration controls
+- Add DramaBox experimental staged and sequential memory strategies, FP8, and optional compilation
+- Add DramaBox near-silence warnings for text and subtitle generation
+- Add ChatterBox 23-Lang V3 checkpoint selection
+
+### Changed
+
+- Improve multiline parameter controls and generated-audio cache accuracy
+- Update engine comparison tables, model download information, and user guides
+## [5.5.3] - 2026-07-24
+
+### Added
+
+- Add an in-ComfyUI Character Alias Manager for creating, organizing, previewing, and overriding character aliases
+- Add Character Alias Manager access from Character Voices and the Multiline TTS Tag Editor
+
+### Changed
+
+- Improve Character Voices waveform clarity, canvas zoom behavior, character discovery, and console logging
+## [5.5.2] - 2026-07-21
+
+### Fixed
+
+- Fix IndexTTS-2 emotion vector importing
+- Fix the Import dialog appearing behind the emotion vector editor
+## [5.5.1] - 2026-07-18
+
+### Added
+
+- Add selectable shared and dedicated runtimes to the Step Audio EditX Engine node
+
+### Changed
+
+- Improve Step Audio EditX memory use and generation reliability
+
+### Fixed
+
+- Fix Step Audio EditX voice cloning producing silence, invalid speech, or assistant-like output
+- Fix Step Audio EditX inline emotion and style editing with isolated runtimes
+- Improve Step Audio EditX progress reporting and compatibility warnings
+## [5.5.0] - 2026-07-17
+
+### Added
+
+- Add MOSS-TTS v1.5 with expanded multilingual speech generation
+- Add MOSS-SoundEffect v1 and MOSS-SoundEffect v2 text-to-sound generation
+- Add unified Voice Designer support for Qwen3-TTS, MOSS-TTS, and OmniVoice
+- Add Save Character Voice for reusable generated or imported voices
+- Add Sound Effects parameter switching, pauses, chunking, crossfades, negative prompts, and audio caching
+- Add MOSS-TTS v1.5 LoRA training support
+- Add Voice Designer and Sound Effects example workflows and user guides
+
+### Changed
+
+- Improve Character Voices discovery, trimming, compact layouts, and immediate saved-voice availability
+- Improve model selection and Hugging Face download progress across supported engines
+
+### Removed
+
+- Remove the legacy Qwen3-TTS Voice Designer node; use Voice Designer instead
+## [5.4.16] - 2026-07-16
+
+### Fixed
+
+- Fix Qwen and Character Voices compatibility
+- Fix Qwen legacy runtimes failing with newer inherited dependencies
+- Fix Qwen text generation stopping after multi-block input
+- Fix old Character Voices workflows loading without their saved voice transcription
+## [5.4.15] - 2026-07-15
+
+### Fixed
+
+- Fix Fish Audio S2 and local VibeVoice loading
+- Fix Fish Audio S2 failing with recent TorchAudio versions
+- Fix local VibeVoice models not being found by Shared or Dedicated Runtime
+## [5.4.14] - 2026-07-15
+
+### Fixed
+
+- Fix engine installation failures reported on Python 3.13
+- Fix Fish Audio S2 failing to load after installation
+- Fix Dots TTS failing when optional text normalization is unavailable
+- Fix VibeVoice Shared Runtime installation failing on Windows
+## [5.4.13] - 2026-07-14
+
+### Added
+
+- Warn that HuBERT Large training is experimental and may produce unintelligible audio
+- Recommend ContentVec 768 for reliable RVC voice training
+
+### Fixed
+
+- Fix RVC voice conversion selecting an incompatible feature encoder
+- Automatically match RVC voice models with the correct feature encoder
+## [5.4.12] - 2026-07-14
+
+### Added
+
+- Allow Dots installation where its Python 3.13 source path works
+
+### Fixed
+
+- Fix Fish Audio S2 and Dots installation on Python 3.13
+- Repair Fish S2 installations missing the inference runtime
+## [5.4.11] - 2026-07-13
+
+### Fixed
+
+- Fix RVC Dataset Prep failing on first runs or incomplete cached datasets
+- Rebuild missing RVC training features automatically instead of stopping on missing feature directory errors
+## [5.4.10] - 2026-07-13
+
+### Added
+
+- Document reference transcript requirements across TTS engines
+- Add a Reference Transcript row to the engine Feature Comparison
+- Clarify which engines require, conditionally use, optionally use, or ignore transcripts
+- Add mode-specific notes for CosyVoice3, Qwen3-TTS, and MOSS-TTS
+## [5.4.9] - 2026-07-13
+
+### Added
+
+- Refine Character Voices waveform and discovery
+- Add a compact normalized waveform to Character Voices trim controls
+- Add playback progress and a smooth playhead within the waveform
+
+### Changed
+
+- Improve trim warning stability without shifting the node layout
+
+### Fixed
+
+- Fix repeated Character Voices discovery scans and console messages
+## [5.4.8] - 2026-07-13
+
+### Added
+
+- Prevent silent output from the unsupported FP16 flow and vocoder path
+
+### Fixed
+
+- Fix CosyVoice3 generation on ROCm systems
+- Fix CosyVoice3 generation failures caused by incompatible mixed precision
+- Show the underlying generation error instead of a misleading follow-on error
+## [5.4.7] - 2026-07-13
+
+### Added
+
+- Add automatic reference transcription loading with live workflow editing
+- Add draggable audio trimming with bounded playback and precise time controls
+- Add a reference-audio-only output for reuse in audio workflows
+
+### Changed
+
+- Enhance Character Voices reference editing
+- Improve customized voice handling across Unified Text and SRT engines
+
+### Fixed
+
+- Fix trimmed character voices using the original untrimmed source in some engines
+## [5.4.6] - 2026-07-13
+
+### Added
+
+- Add an editable import dialog that matches the export interface
+- Let users paste and adjust JSON values before applying them
+- Validate emotion values before updating the node
+- Make import controls clear and consistent with export
+
+### Changed
+
+- Improve IndexTTS-2 emotion vector import
+## [5.4.5] - 2026-07-12
+
+### Added
+
+- Repair incomplete or incompatible Fish Speech installations automatically
+
+### Fixed
+
+- Fix Fish Audio S2 installation on clean environments
+- Prevent Fish S2 generation from failing because its inference runtime is missing
+## [5.4.4] - 2026-07-11
+
+### Added
+
+- Add faster character, parameter, preset, and emotion swapping
+- Support combining IndexTTS audio emotion references with vector or text emotions
+
+### Changed
+
+- Improve Multiline TTS Tag Editor emotion switching
+- Improve long inline tag wrapping inside the editor
+
+### Fixed
+
+- Fix incorrect tag detection and extra blank lines when pressing Enter
+## [5.4.3] - 2026-07-10
+
+### Added
+
+- Keep emotion radar controls aligned at the original node size
+- Prevent the radar chart from overflowing its node
+- Make emotion vector export selectable and downloadable on demand
+- Show confirmation after importing emotion vectors
+
+### Fixed
+
+- Fix IndexTTS-2 emotion vector controls
+## [5.4.2] - 2026-07-10
+
+### Added
+
+- Corrects pitch-index typing for RVC voice conversion on MPS devices.
+- Preserves the model's expected precision for phone features across CPU, CUDA, and MPS.
+
+### Fixed
+
+- Tentative fix for RVC Voice Changer on Apple Silicon
+## [5.4.1] - 2026-07-10
+
+### Added
+
+- Prevents MelBand vocal removal from failing during sample-rate conversion on some Python 3.13 environments.
+- Corrects the documented location of the version bump instructions.
+- Makes future version bumps reliable on Windows installations with non-UTF-8 locales.
+
+### Fixed
+
+- Tentative fix for MelBand audio separation failures
+## [5.4.0] - 2026-07-10
+
+### Added
+
+- Add Fish Audio S2 Pro multilingual voice generation
+- Add Fish Audio S2 Pro voice cloning with reference audio and transcript support
+- Add native multi-speaker dialogue and independent character-segment generation
+- Add free-form inline speech instructions and automatic language prompting
+- Add long-form generation, SRT integration, compilation, caching, and optional quantization
+## [5.3.0] - 2026-06-23
+
+### Added
+
+- Add Granite ASR 4.1 diarization and timestamp improvements
+- Add Granite ASR plus-model speaker diarization with suite-native [Speaker] output for downstream TTS workflows
+- Clarify Granite ASR 4.1 model differences, diarization support, and Japanese support in the README and generated docs
+
+### Changed
+
+- Improve test runner portability and Granite ASR regression coverage for contributors
+
+### Fixed
+
+- Fix Granite ASR plus native word timestamps cutting off longer transcripts in timestamp mode
+## [5.2.0] - 2026-06-23
+
+### Added
+
+- Add OmniVoice engine with native timing-aware TTS controls
+- Add OmniVoice text and SRT generation with native duration targeting for more precise subtitle timing
+- Add Visual Tag Builder workflow example and documentation for OmniVoice prompt building
+
+### Changed
+
+- Improve interruption handling so stopped generation exits cleanly across unified TTS workflows
+
+### Fixed
+
+- Fix Qwen3 Voice Designer metadata links and test environment path handling
+## [5.1.1] - 2026-06-17
+
+### Added
+
+- Add a virtualenv fallback when the bundled Python cannot create isolated runtimes with venv
+- Keep isolated runtimes reusing the main ComfyUI Torch and other heavy dependencies when inheritance succeeds
+
+### Changed
+
+- Improve runtime bootstrap errors so users can see why environment creation failed
+
+### Fixed
+
+- Fix Unified ASR isolated runtime setup on portable ComfyUI
+- Fix Qwen3 ASR and Granite forced-aligner workflows failing before model startup on some portable Windows installs
+## [5.1.0] - 2026-06-15
+
+### Added
+
+- Add Dots TTS engine with multilingual cloning support
+- Add official Dots TTS engine for text and SRT generation
+- Refine language behavior, Portuguese metadata, and Dots engine documentation
+
+### Changed
+
+- Improve model unload, caching, and generation progress feedback
+## [5.0.0] - 2026-06-14
+
+### Added
+
+- Add Higgs Audio v3 TTS engine with native inline tags and zero-shot voice cloning
+- Add isolated runtimes for fragile legacy engines so the main environment no longer has to stay on the older stack
+
+### Changed
+
+- Improve the Multiline TTS Tag Editor with engine-aware inline tags, validation, conversion, and persistence
+- Update workflows, screenshots, and documentation for the new architecture
+## [4.27.3] - 2026-05-26
+
+### Added
+
+- Default RVC voice conversion to native long-audio segmentation
+- Add an RVC option for extra outer chunking when lower-VRAM chunk sizes are needed
+- Clarify which chunking mode is active in RVC conversion info
+
+### Changed
+
+- Change RVC custom chunking to opt-in
+## [4.27.2] - 2026-05-26
+
+### Added
+
+- Preserve duration more accurately for long audio and training workflows
+
+### Fixed
+
+- Fix RVC long-audio chunk timing drift
+- Fix RVC voice conversion outputs getting shorter over long chunked runs
+- Keep the fix inside the bundled RVC pipeline instead of stretching audio afterward
+## [4.27.1] - 2026-05-26
+
+### Added
+
+- Preserve timing when rejoining chunked audio across all VC engines
+
+### Changed
+
+- Clarify smart chunking guidance for cleaner boundaries
+
+### Fixed
+
+- Fix Unified Voice Changer long-audio chunk timing drift
+- Fix chunked voice conversion shortening long outputs over time
 ## [4.27.0] - 2026-05-26
 
 ### Added

LICENSE

@@ -27,22 +27,28 @@ Third-Party Model Licenses
 
 The project code is MIT. Model weights carry their own licenses:
 
-  ────────────────  ─────────────────────────────────────────  ──────────────
-  Engine            License                                    Commercial Use
-  ────────────────  ─────────────────────────────────────────  ──────────────
-  F5-TTS            CC-BY-NC-4.0                               No            
-  ChatterBox        MIT                                        Yes           
-  ChatterBox 23L    MIT                                        Yes           
-  VibeVoice         MIT (research-only per model card)         No            
-  Higgs Audio 2     Boson Higgs Audio 2 Community License      Conditional   
-  IndexTTS-2        bilibili Model Use License                 Conditional   
-  CosyVoice3        Apache-2.0                                 Yes           
-  Qwen3-TTS         Apache-2.0                                 Yes           
-  Granite ASR       Apache-2.0                                 Yes           
-  Step Audio EditX  Apache-2.0 (verify before commercial use)  Conditional   
-  Echo-TTS          CC-BY-NC-SA-4.0                            No            
-  MOSS-TTS          Apache-2.0                                 Yes           
-  RVC               MIT (framework); community models vary     Varies        
-  ────────────────  ─────────────────────────────────────────  ──────────────
+  ───────────────────  ────────────────────────────────────────────────────────  ──────────────
+  Engine               License                                                   Commercial Use
+  ───────────────────  ────────────────────────────────────────────────────────  ──────────────
+  F5-TTS               CC-BY-NC-4.0                                              No
+  ChatterBox           MIT                                                       Yes
+  ChatterBox 23L       MIT                                                       Yes
+  VibeVoice            MIT (research-only per model card)                        No
+  Higgs Audio 2        Boson Higgs Audio 2 Community License                     Conditional
+  Higgs Audio v3       Boson Higgs Audio v3 Research and Non-Commercial License  No
+  IndexTTS 2 / 2.5     bilibili Model Use License                                Conditional
+  CosyVoice3           Apache-2.0                                                Yes
+  Qwen3-TTS            Apache-2.0                                                Yes
+  Granite ASR          Apache-2.0                                                Yes
+  Step Audio EditX     Apache-2.0 (verify before commercial use)                 Conditional
+  Echo-TTS             CC-BY-NC-SA-4.0                                           No
+  Fish Audio S2 Pro    Fish Audio Research License                               No
+  Dots TTS             Apache-2.0                                                Yes
+  DramaBox             LTX-2 Community License                                   Conditional
+  OmniVoice            Apache-2.0                                                Yes
+  MOSS-TTS             Apache-2.0                                                Yes
+  MOSS-SoundEffect v2  Apache-2.0                                                Yes
+  RVC                  MIT (framework); community models vary                    Varies
+  ───────────────────  ────────────────────────────────────────────────────────  ──────────────
 
 Users are responsible for complying with respective model licenses.

PROJECT_INDEX.md

@@ -4,33 +4,37 @@
 
 ## Architecture Overview
 
-**Unified modular architecture** - all engines share the same layered pattern:
-
-1. **Engine Node** (`nodes/engines/<engine>_engine_node.py`) - ComfyUI configuration UI
-2. **Unified Nodes** (`nodes/unified/`) - thin delegation layer, routes to engine processors
-3. **Processor** (`nodes/<engine>/<engine>_processor.py`) - orchestration, chunking, character/pause/tag handling
-4. **SRT Processor** (`nodes/<engine>/<engine>_srt_processor.py`) - SRT-specific orchestration
-5. **Adapter** (`engines/adapters/<engine>_adapter.py`) - bridges processor to engine implementation
-6. **Engine Implementation** (`engines/<engine>/`) - actual model inference
-
-**Key architectural rules:**
-- Chunking happens in the **processor**, not the adapter (`generate_single()` on adapter = raw single call)
+**Unified modular architecture** - all engines share the same layered pattern:
+
+1. **Engine Node** (`nodes/engines/<engine>_engine_node.py`) - ComfyUI configuration UI
+2. **Unified Nodes** (`nodes/unified/`) - thin delegation layer, routes to engine processors
+3. **Processor** (`nodes/<engine>/<engine>_processor.py`) - orchestration, chunking, character/pause/tag handling
+4. **SRT Processor** (`nodes/<engine>/<engine>_srt_processor.py`) - SRT-specific orchestration
+5. **Adapter** (`engines/adapters/<engine>_adapter.py`) - bridges processor to engine implementation
+6. **Engine Implementation** (`engines/<engine>/`) - actual model inference
+7. **Optional Isolated Runtime** (`utils/runtimes/`) - worker subprocess layer for fragile engine stacks
+
+**Key architectural rules:**
+- Chunking happens in the **processor**, not the adapter (`generate_single()` on adapter = raw single call)
+- Runtime routing happens through `ModelLoadConfig.runtime_mode` + `runtime_profile`, not ad-hoc subprocess calls
+- Shared runtime workers are currently used for fragile engine families such as VibeVoice, Qwen3-TTS / ASR, Granite forced alignment, and Higgs Audio 2. Engines that support the modern stack run natively in the main Transformers 5 environment.
 - YAML (`docs/Dev reports/tts_audio_suite_engines.yaml`) is source of truth for engine doc tables → run `python3 scripts/generate_engine_tables.py --readme` to regenerate
 - Auxiliary YAML (`docs/Dev reports/tts_audio_suite_aux_models.yaml`) is source of truth for helper/post-process model docs → run `python3 scripts/generate_aux_model_docs.py`
-- All models download to `ComfyUI/models/TTS/<model-name>/`
-- Engine registry: `utils/models/engine_registry.py`
+- All models download to `ComfyUI/models/TTS/<model-name>/`
+- Engine registry: `utils/models/engine_registry.py`
 
 ## Engines
 
-13 engines follow the pattern above:
+19 engines follow the pattern above:
 
 | Engine | Adapter | Processor | SRT Processor | Engine Node |
 |--------|---------|-----------|---------------|-------------|
-| ChatterBox | `chatterbox_adapter.py` | `nodes/chatterbox/chatterbox_tts_node.py` | `chatterbox_srt_node.py` | `chatterbox_engine_node.py` |
-| ChatterBox 23-Lang | `chatterbox_streaming_adapter.py` | `nodes/chatterbox_official_23lang/` | same folder | `chatterbox_engine_node.py` |
-| F5-TTS | `f5tts_adapter.py` | `nodes/f5tts/f5tts_node.py` | `f5tts_srt_node.py` | `f5tts_engine_node.py` |
-| Higgs Audio 2 | `higgs_audio_adapter.py` | — | `nodes/higgs_audio/higgs_audio_srt_processor.py` | `higgs_audio_engine_node.py` |
-| VibeVoice | `vibevoice_adapter.py` | `nodes/vibevoice/vibevoice_processor.py` | — | `vibevoice_engine_node.py` |
+| ChatterBox | `chatterbox_adapter.py` | `nodes/chatterbox/chatterbox_tts_node.py` | `chatterbox_srt_node.py` | `chatterbox_engine_node.py` |
+| ChatterBox 23-Lang | `chatterbox_official_23lang_adapter.py` | `nodes/chatterbox_official_23lang/` | same folder | `chatterbox_official_23lang_engine_node.py` |
+| F5-TTS | `f5tts_adapter.py` | `nodes/f5tts/f5tts_node.py` | `f5tts_srt_node.py` | `f5tts_engine_node.py` |
+| Higgs Audio 2 | `higgs_audio_adapter.py` | — | `nodes/higgs_audio/higgs_audio_srt_processor.py` | `higgs_audio_engine_node.py` |
+| Higgs Audio v3 | `higgs_audio_v3_adapter.py` | `nodes/higgs_audio_v3/higgs_audio_v3_processor.py` | `higgs_audio_v3_srt_processor.py` | `higgs_audio_v3_engine_node.py` |
+| VibeVoice | `vibevoice_adapter.py` | `nodes/vibevoice/vibevoice_processor.py` | — | `vibevoice_engine_node.py` |
 | IndexTTS-2 | — | `engines/index_tts/` | — | `index_tts_engine_node.py` |
 | Step Audio EditX | `step_audio_editx_adapter.py` | `nodes/step_audio_editx/step_audio_editx_processor.py` | `step_audio_editx_srt_processor.py` | `step_audio_editx_engine_node.py` |
 | CosyVoice3 | `cosyvoice_adapter.py` | `engines/processors/cosyvoice_processor.py` | `nodes/cosyvoice/cosyvoice_srt_processor.py` | `cosyvoice_engine_node.py` |
@@ -38,25 +42,34 @@
 | MOSS-TTS | `moss_tts_adapter.py` | `nodes/moss_tts/moss_tts_processor.py` | `moss_tts_srt_processor.py` | `moss_tts_engine_node.py` |
 | Granite ASR | `asr_granite_adapter.py` | — | — | `granite_asr_engine_node.py` |
 | Echo-TTS | `echo_tts_adapter.py` | `nodes/echo_tts/echo_tts_processor.py` | `echo_tts_srt_processor.py` | `echo_tts_engine_node.py` |
+| Fish Audio S2 Pro | `fish_audio_s2_adapter.py` | `nodes/fish_audio_s2/fish_audio_s2_processor.py` | `fish_audio_s2_srt_processor.py` | `fish_audio_s2_engine_node.py` |
+| Dots TTS | `dots_tts_adapter.py` | `nodes/dots_tts/dots_tts_processor.py` | `dots_tts_srt_processor.py` | `dots_tts_engine_node.py` |
+| DramaBox | `dramabox_adapter.py` | `nodes/dramabox/dramabox_processor.py` | `dramabox_srt_processor.py` | `dramabox_engine_node.py` |
+| OmniVoice | `omnivoice_adapter.py` | `nodes/omnivoice/omnivoice_processor.py` | `omnivoice_srt_processor.py` | `omnivoice_engine_node.py` |
+| MOSS-SoundEffect v2 | `moss_soundeffect_v2_adapter.py` | — | — | `moss_soundeffect_v2_engine_node.py` |
 | RVC | — | `engines/rvc/` | — | `rvc_engine_node.py` |
 
-**Engine implementations live in:**
-- `engines/chatterbox/`, `engines/chatterbox_official_23lang/`, `engines/f5tts/`, `engines/higgs_audio/`, `engines/vibevoice_engine/`, `engines/step_audio_editx/`, `engines/cosyvoice/`, `engines/qwen3_tts/`, `engines/qwen3_asr/`, `engines/moss_tts/`, `engines/granite_asr/`, `engines/rvc/`
+**Engine implementations live in:**
+- `engines/chatterbox/`, `engines/chatterbox_official_23lang/`, `engines/f5tts/`, `engines/higgs_audio/`, `engines/higgs_audio_v3/`, `engines/vibevoice_engine/`, `engines/step_audio_editx/`, `engines/cosyvoice/`, `engines/qwen3_tts/`, `engines/qwen3_asr/`, `engines/moss_tts/`, `engines/moss_soundeffect_v2/`, `engines/granite_asr/`, `engines/echo_tts/`, `engines/fish_audio_s2/`, `engines/dots_tts/`, `engines/dramabox/`, `engines/omnivoice/`, `engines/rvc/`
 
 ## Documentation Files
 
-**README.md** - Main project docs, installation, features overview
-**CLAUDE.md** - Dev guidelines for Claude Code
-**CHANGELOG.md** - Full version history
+**README.md** - Main project docs, installation, features overview
+**CLAUDE.md** - Dev guidelines for Claude Code
+**CHANGELOG.md** - Full version history
+**docs/BUMP_SCRIPT_INSTRUCTIONS.md** - Version bump process
 
 ### User Docs (`docs/`)
 - `CHARACTER_SWITCHING_GUIDE.md` - [CharacterName] tag system
 - `PARAMETER_SWITCHING_GUIDE.md` - Per-segment parameter override syntax
 - `INLINE_EDIT_TAGS_USER_GUIDE.md` - Step Audio EditX inline tags
+- `HIGGS_AUDIO_V3_INLINE_TAGS.md` - Higgs Audio v3 native paralinguistic tags
+- `OMNIVOICE_TAGS_GUIDE.md` - OmniVoice native non-verbal tags and pronunciation overrides
 - `MOSS_TTS_PROMPT_FIELDS_GUIDE.md` - Official MOSS whole-segment prompt fields and inline `<>` translation limits
+- `DRAMABOX_PROMPTING_GUIDE.md` - DramaBox expressive scene prompts, voice references, controls, hardware, and license
 - `COSYVOICE3_TAGS_GUIDE.md` - CosyVoice3 native paralinguistic tags
 - `CHATTERBOX_V2_SPECIAL_TOKENS.md` - ChatterBox v2 emotion tokens
-- `IndexTTS2_Emotion_Control_Guide.md` - IndexTTS-2 emotion vectors
+- `IndexTTS2_Emotion_Control_Guide.md` - IndexTTS-2 vector, text, audio, and blended emotion controls
 - `VOCAL_REMOVAL_GUIDE.md` - Vocal separation guide
 - `qwen3_tts_optimizations.md` - Qwen3-TTS torch.compile setup
 - `MODEL_DOWNLOAD_SOURCES.md` - All HF repo links (auto-generated)
@@ -64,11 +77,13 @@
 - `AUX_MODEL_SOURCES.md`, `AUX_MODEL_LAYOUTS.md` - Auxiliary helper-model registries (auto-generated)
 - `ENGINE_COMPARISON.md`, `LANGUAGE_SUPPORT.md`, `FEATURE_COMPARISON.md` - Auto-generated tables
 
-### Dev Docs (`docs/Dev reports/`)
+### Dev Docs (`docs/Dev reports/`)
 - `tts_audio_suite_engines.yaml` - **Source of truth** for all engine metadata
 - `tts_audio_suite_aux_models.yaml` - **Source of truth** for helper/post-process model metadata
-- `BUMP_SCRIPT_INSTRUCTIONS.md` - Version bump process
 - `SRT_IMPLEMENTATION.md` - SRT timing technical details
+- `ISOLATED_RUNTIMES_PLAN.md` - original runtime isolation plan and scope
+- `TRANSFORMERS_5_QWEN3_TTS_REPORT.md` - why Qwen3-TTS moved to shared legacy T4 runtime
+- `TRANSFORMERS_5_HIGGS_AUDIO_REPORT.md` - why Higgs Audio 2 moved to shared legacy T4 runtime
 
 ### New Engine Guides (`docs/New Engines Guides/`)
 - `README.md` - Start-here friendly workflow for users guiding LLMs through new engine integrations
@@ -88,14 +103,17 @@
 
 ## Node Implementations
 
-### Unified Interface
-- `nodes/unified/tts_text_node.py` - Universal TTS text node
-- `nodes/unified/tts_srt_node.py` - Universal SRT node
-- `nodes/unified/voice_changer_node.py` - Universal voice conversion
-- `nodes/unified/asr_transcribe_node.py` - Universal ASR node
+### Unified Interface
+- `nodes/unified/tts_text_node.py` - Universal TTS text node
+- `nodes/unified/tts_srt_node.py` - Universal SRT node
+- `nodes/unified/voice_changer_node.py` - Universal voice conversion
+- `nodes/unified/asr_transcribe_node.py` - Universal ASR node
 
-### Shared / Special Nodes
+### Shared / Special Nodes
 - `nodes/shared/character_voices_node.py` - Character voice management (NARRATOR_VOICE output)
+- `nodes/shared/unified_voice_designer_node.py` - Unified Qwen VoiceDesign, MOSS VoiceGenerator, and reference-free OmniVoice design
+- `nodes/shared/save_character_voice_node.py` - Explicit output node for saving any NARRATOR_VOICE into the established voice library
+- `nodes/omnivoice/omnivoice_instruction_builder_node.py` - OmniVoice voice-design instruction helper with custom visual builder UI
 - `nodes/text/phoneme_text_normalizer_node.py` - Multilingual text preprocessing
 - `nodes/text/asr_punctuation_truecase_node.py` - Standalone punctuation / truecase cleanup for raw ASR text
 - `nodes/subtitles/text_to_srt_builder_node.py` - Build SRT from transcript text plus timing data, or estimate timings from plain text
@@ -103,7 +121,6 @@
 - `nodes/text/tts_tag_editor_node.py` - 🏷️ Multiline TTS Tag Editor: rich text editor with character/language/parameter dropdowns, preset system, syntax highlighting, undo/redo — pairs with `web/string_multiline_tag_editor.js`
 - `nodes/step_audio_editx_special/step_audio_editx_audio_editor_node.py` - 🎨 Audio Editor: post-process ANY engine's audio with Step Audio EditX (14 emotions, 32 styles, paralinguistic effects like `<Laughter>`, speed control) — universal, not just for Step Audio EditX engine
 - `nodes/engines/index_tts_emotion_options_node.py` - IndexTTS-2 emotion radar chart
-- `nodes/qwen3_tts/qwen3_tts_voice_designer_node.py` - Qwen3 voice-from-text-description
 
 ### Audio / Video Nodes
 - `nodes/audio/analyzer_node.py` - Audio Wave Analyzer
@@ -115,12 +132,26 @@
 
 ## Utility Systems
 
-### Model Management (`utils/models/`)
-- `unified_model_interface.py` - Universal factory pattern for all engines
-- `engine_registry.py` - Engine capability definitions
-- `manager.py` - Model discovery and caching
-- `comfyui_model_wrapper/` - ComfyUI native model management integration
-- `extra_paths.py` - extra_model_paths.yaml support
+### Model Management (`utils/models/`)
+- `unified_model_interface.py` - Universal factory pattern for all engines
+- `engine_registry.py` - Engine capability definitions
+- `factory_config.py` - standardized model load config, runtime mode/profile normalization
+- `manager.py` - Model discovery and caching
+- `comfyui_model_wrapper/` - ComfyUI native model management integration
+- `extra_paths.py` - extra_model_paths.yaml support
+
+### Isolated Runtimes (`utils/runtimes/`)
+- `profiles.py` - named runtime profiles (`vibevoice_transformers4_shared`, dedicated variants, etc.)
+- `launcher.py` - runtime bootstrap, venv creation, Windows toolchain env setup
+- `session.py`, `protocol.py` - JSONL worker transport and message protocol
+- `bootstrap.py` - shared runtime bootstrap helpers
+- `vibevoice_proxy.py`, `qwen3_tts_proxy.py`, `qwen3_asr_proxy.py`, `higgs_audio_proxy.py` - parent-process proxies
+- `workers/` - worker subprocess entrypoints for VibeVoice, Qwen3-TTS, Qwen3-ASR/aligner, Higgs Audio
+- Current shared legacy T4 runtime profile is reused by:
+  - VibeVoice / Kugel
+  - Qwen3-TTS
+  - Qwen3-ASR and Granite's optional Qwen forced aligner
+  - Higgs Audio 2
 
 ### Audio (`utils/audio/`)
 - `processing.py` - Tensor manipulation, normalization, format conversion
@@ -146,9 +177,12 @@
 - `parser.py` - SRT parsing and validation
 - `reporting.py` - Timing report generation
 
-### Other Utils
-- `utils/voice/discovery.py` - Voice file discovery with multi-path fallback
-- `utils/downloads/unified_downloader.py` - Centralized HF download system
+### Other Utils
+- `utils/voice/discovery.py` - Voice discovery, user-voice priority, and multi-path fallback
+- `utils/voice/designers.py` - Whitelisted voice-designer provider registry
+- `utils/voice/character_saver.py` - Shared `.wav` / `.reference.txt` / `.txt` character persistence
+- `utils/voice/character_logging.py` - Shared resolved voice labels and boxed prompt previews
+- `utils/downloads/unified_downloader.py` - Centralized HF download system
 - `utils/compatibility/transformers_patches.py` - transformers version compatibility patches
 - `utils/compatibility/numba_compat.py` - Numba/Librosa Python 3.13+ compatibility
 - `utils/ffmpeg_utils.py` - FFmpeg with graceful fallback
@@ -162,11 +196,11 @@
 ### Audio Analyzer
 `web/audio_analyzer_*.js` (core, ui, visualization, regions, controls, widgets, drawing, events, layout, node_integration)
 
-### Other Web Files
-- `web/chatterbox_voice_capture.js` - Microphone recording UI
-- `web/index_tts_emotion_radar.js` + `emotion_radar_canvas_widget.js` - IndexTTS-2 radar chart
-- `web/qwen3_tts_widgets.js` - Qwen3 conditional instruction field
-- `web/asr_srt_preset_widgets.js` - ASR SRT preset locking
+### Other Web Files
+- `web/chatterbox_voice_capture.js` - Microphone recording UI
+- `web/index_tts_emotion_radar.js` + `emotion_radar_canvas_widget.js` - IndexTTS-2 radar chart
+- `web/qwen3_tts_widgets.js` - Qwen model-specific widget enablement and legacy workflow migration
+- `web/asr_srt_preset_widgets.js` - ASR SRT preset locking
 
 ## Scripts & Config
 - `scripts/bump_version_enhanced.py` - Version bump with changelog (use `patch`/`minor`/`major`)

__init__.py

@@ -12,9 +12,51 @@ Unified architecture supporting ChatterBox, F5-TTS, and future engines like RVC:
 # Setting it here causes "allocator mismatch" errors because ComfyUI already imported torch
 
 # Import from the main nodes.py file which handles the new unified architecture
-import importlib.util
-import os
-import sys
+import importlib.util
+import os
+import sys
+
+# ComfyUI 0.12+ owns a top-level ``utils`` package, while this long-standing
+# node pack also imports its helpers through ``utils.*``. Preserve ComfyUI's
+# loaded package and extend only its module search path with this pack's utils.
+_project_root = os.path.dirname(__file__)
+_suite_utils_root = os.path.abspath(os.path.join(_project_root, "utils"))
+if _project_root in sys.path:
+    sys.path.remove(_project_root)
+sys.path.insert(0, _project_root)
+
+_loaded_utils = sys.modules.get("utils")
+if _loaded_utils is None:
+    import utils as _loaded_utils
+
+_utils_search_path = getattr(_loaded_utils, "__path__", None)
+if _utils_search_path is None:
+    raise ImportError(
+        "TTS Audio Suite cannot extend the loaded top-level 'utils' module because it is not a package"
+    )
+
+_normalized_utils_paths = {os.path.normcase(os.path.abspath(path)) for path in _utils_search_path}
+if os.path.normcase(_suite_utils_root) not in _normalized_utils_paths:
+    _utils_search_path.insert(0, _suite_utils_root)
+
+# When this pack is imported before ComfyUI imports its own helpers, locate the
+# active ComfyUI utils directory by its stable core modules and add it as the
+# fallback side of the same package search path.
+for _search_root in sys.path:
+    _candidate_utils = os.path.abspath(os.path.join(_search_root or os.curdir, "utils"))
+    _normalized_candidate = os.path.normcase(_candidate_utils)
+    if _normalized_candidate in _normalized_utils_paths or _normalized_candidate == os.path.normcase(_suite_utils_root):
+        continue
+    if all(os.path.isfile(os.path.join(_candidate_utils, filename)) for filename in ("extra_config.py", "install_util.py")):
+        _utils_search_path.append(_candidate_utils)
+        _normalized_utils_paths.add(_normalized_candidate)
+
+from utils.hf_download_logging import configure_hf_download_logging
+
+
+# Keep every engine's Hugging Face download output readable. Download failures
+# are still reported by the suite's downloader error handling.
+configure_hf_download_logging()
 
 # Note: PyTorch inductor patches removed - not needed for PyTorch 2.10+ with triton-windows 3.6+
 # Qwen3-TTS torch.compile optimizations require:
@@ -114,8 +156,9 @@ def check_dependencies():
         print(f"{'='*80}")
         print(f"The following required packages are missing: {', '.join(missing)}")
         print(f"")
-        print(f"Please run the installation script or install them manually:")
-        print(f"pip install -r requirements.txt")
+        install_script = os.path.join(os.path.dirname(__file__), "install.py")
+        print(f"Please run the TTS Audio Suite installation script:")
+        print(f'"{sys.executable}" "{install_script}"')
         print(f"{'='*80}\n")
 
 # Version disclosure for troubleshooting
@@ -149,33 +192,7 @@ def print_critical_versions():
 
     print(f"ℹ️ Critical package versions: {', '.join(version_info)}")
 
-def warn_transformers_5_unsupported():
-    """Warn when Transformers 5.x is installed (Qwen3-TTS tokenizer is incompatible).
-
-    NOTE: This check uses sys.modules to avoid eagerly importing transformers (~1.3s).
-    If transformers hasn't been imported yet (e.g. by the version-printing function above),
-    we skip the check -- it will be caught later when an engine actually loads transformers.
-    """
-    try:
-        # Only check if transformers is already loaded (avoids ~1.3s eager import)
-        import sys as _sys
-        if 'transformers' not in _sys.modules:
-            return
-        transformers = _sys.modules['transformers']
-        try:
-            from packaging.version import Version
-            version = Version(transformers.__version__)
-            is_5x = version >= Version("5.0.0")
-        except Exception:
-            parts = transformers.__version__.split(".")
-            is_5x = int(parts[0]) >= 5 if parts and parts[0].isdigit() else False
-        if is_5x:
-            print("⚠️ Transformers 5.x detected: Qwen3-TTS tokenizer is incompatible.")
-            print("   Please downgrade to transformers<=4.57.3 (see requirements.txt).")
-    except Exception:
-        pass
-
-def check_ffmpeg_availability():
+def check_ffmpeg_availability():
     """Check ffmpeg availability and log status"""
     try:
         # Load ffmpeg_utils directly by file path to avoid package import issues
@@ -204,10 +221,9 @@ def check_ffmpeg_availability():
             print("⚠️ FFmpeg not found - using fallback audio processing (reduced quality)")
             print("💡 Install FFmpeg for optimal performance: https://ffmpeg.org/download.html")
 
-# Print versions and check dependencies immediately for troubleshooting
-check_dependencies()
-print_critical_versions()
-warn_transformers_5_unsupported()
+# Print versions and check dependencies immediately for troubleshooting
+check_dependencies()
+print_critical_versions()
 check_ffmpeg_availability()
 
 # Check for old ChatterBox extension conflict
@@ -321,35 +337,68 @@ __all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS"]
 WEB_DIRECTORY = "./web"
 
 # Register API endpoint for widget data
-def setup_api_routes():
-    """Setup API routes for widget communication"""
-    try:
-        import json
-        from server import PromptServer
-        from aiohttp import web
+def setup_api_routes():
+    """Setup API routes for widget communication"""
+    try:
+        import json
+        import folder_paths
+        from server import PromptServer
+        from aiohttp import web
+
+        def _get_ui_data_dir():
+            base_dir = os.path.join(folder_paths.get_system_user_directory("tts_audio_suite"), "ui")
+            os.makedirs(base_dir, exist_ok=True)
+            return base_dir
+
+        def _get_omnivoice_preset_library_path():
+            return os.path.join(_get_ui_data_dir(), "omnivoice_instruction_builder_presets.json")
+
+        from utils.voice.alias_api import register_character_alias_routes
+        register_character_alias_routes(PromptServer.instance.routes, web)
+        from utils.audio_cpp.capability_api import register_audio_cpp_capability_routes
+        register_audio_cpp_capability_routes(PromptServer.instance.routes, web)
+
+        @PromptServer.instance.routes.get("/api/tts-audio-suite/index-tts-emotion-presets")
+        async def get_index_tts_emotion_presets_endpoint(request):
+            """Return presets stored beside the IndexTTS resources under models/TTS."""
+            try:
+                from .utils.text.index_tts_emotion import load_emotion_presets
+                return web.json_response({"presets": load_emotion_presets()})
+            except Exception as e:
+                print(f"⚠️ Error retrieving IndexTTS emotion presets: {e}")
+                return web.json_response({"presets": {}, "error": str(e)}, status=500)
+
+        @PromptServer.instance.routes.post("/api/tts-audio-suite/index-tts-emotion-presets")
+        async def save_index_tts_emotion_presets_endpoint(request):
+            """Atomically persist the IndexTTS emotion preset library."""
+            try:
+                from .utils.text.index_tts_emotion import save_emotion_presets
+                data = await request.json()
+                presets = data.get("presets", {})
+                path = save_emotion_presets(presets)
+                return web.json_response({"status": "success", "count": len(presets), "path": path})
+            except ValueError as e:
+                return web.json_response({"error": str(e)}, status=400)
+            except Exception as e:
+                print(f"⚠️ Error saving IndexTTS emotion presets: {e}")
+                return web.json_response({"status": "error", "error": str(e)}, status=500)
 
-        @PromptServer.instance.routes.get("/api/tts-audio-suite/available-characters")
-        async def get_available_characters_endpoint(request):
-            """API endpoint to get available TTS character voices including aliases"""
-            try:
-                # Load voice discovery directly by file path to avoid package import issues
-                voice_discovery_path = os.path.join(os.path.dirname(__file__), "utils", "voice", "discovery.py")
-                spec = importlib.util.spec_from_file_location("voice_discovery_module", voice_discovery_path)
-                voice_discovery_module = importlib.util.module_from_spec(spec)
-                spec.loader.exec_module(voice_discovery_module)
-
-                characters = list(voice_discovery_module.get_available_characters())
-                # Also get character aliases
-                aliases = list(voice_discovery_module.voice_discovery._character_aliases.keys()) if hasattr(voice_discovery_module.voice_discovery, '_character_aliases') else []
-                # Combine and deduplicate
-                all_chars = sorted(set(characters + aliases))
+        @PromptServer.instance.routes.get("/api/tts-audio-suite/available-characters")
+        async def get_available_characters_endpoint(request):
+            """API endpoint to get available TTS character voices including aliases"""
+            try:
+                from utils.voice import discovery as voice_discovery_module
+                characters = list(voice_discovery_module.get_available_characters())
+                aliases = list(voice_discovery_module.voice_discovery.get_character_aliases().keys())
+                # Combine and deduplicate
+                all_chars = sorted(set(characters + aliases))
                 return web.json_response({"characters": all_chars})
             except Exception as e:
                 print(f"⚠️ Error retrieving available characters: {e}")
                 return web.json_response({"characters": [], "error": str(e)})
 
-        @PromptServer.instance.routes.get("/api/tts-audio-suite/available-languages")
-        async def get_available_languages_endpoint(request):
+        @PromptServer.instance.routes.get("/api/tts-audio-suite/available-languages")
+        async def get_available_languages_endpoint(request):
             """API endpoint to get available language codes from the canonical language mapper"""
             try:
                 # Load language_mapper directly by file path to avoid package import issues
@@ -362,12 +411,59 @@ def setup_api_routes():
                 languages = sorted(set(language_mapper_module.LANGUAGE_ALIASES.values()))
                 return web.json_response({"languages": languages})
             except Exception as e:
-                print(f"⚠️ Error retrieving available languages: {e}")
-                # Fallback list
-                return web.json_response({"languages": ["en", "de", "fr", "ja", "es", "it", "pt", "th", "no"], "error": str(e)})
-
-        @PromptServer.instance.routes.get("/api/tts-audio-suite/voice-input-devices")
-        async def get_voice_input_devices_endpoint(request):
+                print(f"⚠️ Error retrieving available languages: {e}")
+                # Fallback list
+                return web.json_response({"languages": ["en", "de", "fr", "ja", "es", "it", "pt", "th", "no"], "error": str(e)})
+
+        @PromptServer.instance.routes.get("/api/tts-audio-suite/omnivoice-presets")
+        async def get_omnivoice_presets_endpoint(request):
+            """Return the persisted OmniVoice instruction builder preset library."""
+            try:
+                library_path = _get_omnivoice_preset_library_path()
+                if not os.path.exists(library_path):
+                    return web.json_response({"presets": [], "builtinStates": {}, "builtinLayouts": {}})
+                with open(library_path, "r", encoding="utf-8") as f:
+                    payload = json.load(f)
+                presets = payload.get("presets", []) if isinstance(payload, dict) else []
+                builtin_states = payload.get("builtinStates", {}) if isinstance(payload, dict) else {}
+                builtin_layouts = payload.get("builtinLayouts", {}) if isinstance(payload, dict) else {}
+                if not isinstance(presets, list):
+                    presets = []
+                if not isinstance(builtin_states, dict):
+                    builtin_states = {}
+                if not isinstance(builtin_layouts, dict):
+                    builtin_layouts = {}
+                return web.json_response({"presets": presets, "builtinStates": builtin_states, "builtinLayouts": builtin_layouts})
+            except Exception as e:
+                print(f"⚠️ Error retrieving OmniVoice preset library: {e}")
+                return web.json_response({"presets": [], "builtinStates": {}, "builtinLayouts": {}, "error": str(e)}, status=500)
+
+        @PromptServer.instance.routes.post("/api/tts-audio-suite/omnivoice-presets")
+        async def save_omnivoice_presets_endpoint(request):
+            """Persist the OmniVoice instruction builder preset library."""
+            try:
+                data = await request.json()
+                presets = data.get("presets", [])
+                builtin_states = data.get("builtinStates", {})
+                builtin_layouts = data.get("builtinLayouts", {})
+                if not isinstance(presets, list):
+                    return web.json_response({"error": "presets must be a list"}, status=400)
+                if not isinstance(builtin_states, dict):
+                    return web.json_response({"error": "builtinStates must be an object"}, status=400)
+                if not isinstance(builtin_layouts, dict):
+                    return web.json_response({"error": "builtinLayouts must be an object"}, status=400)
+
+                library_path = _get_omnivoice_preset_library_path()
+                payload = {"presets": presets, "builtinStates": builtin_states, "builtinLayouts": builtin_layouts}
+                with open(library_path, "w", encoding="utf-8") as f:
+                    json.dump(payload, f, ensure_ascii=False, indent=2)
+                return web.json_response({"status": "success", "count": len(presets)})
+            except Exception as e:
+                print(f"⚠️ Error saving OmniVoice preset library: {e}")
+                return web.json_response({"status": "error", "error": str(e)}, status=500)
+
+        @PromptServer.instance.routes.get("/api/tts-audio-suite/voice-input-devices")
+        async def get_voice_input_devices_endpoint(request):
             """Return input devices without risking a main-process PortAudio hang."""
             try:
                 import json
@@ -460,8 +556,33 @@ print(json.dumps({"devices": devices}))
                 print(f"⚠️ Error setting inline tag settings: {e}")
                 return web.json_response({"status": "error", "error": str(e)})
 
-        @PromptServer.instance.routes.get("/api/tts-audio-suite/voice-preview")
-        async def get_voice_preview_endpoint(request):
+        def get_voice_discovery_module():
+            """Return the shared discovery module used by nodes and save notifications."""
+            from utils.voice import discovery as voice_discovery_module
+            return voice_discovery_module
+
+        def resolve_character_voice(voice_name):
+            """Resolve a dropdown key through the shared discovery cache."""
+            voice_discovery_module = get_voice_discovery_module()
+            voice_discovery_module.get_available_voices(force_refresh=False)
+            return voice_discovery_module.load_voice_reference(voice_name)
+
+        @PromptServer.instance.routes.get("/api/tts-audio-suite/voice-library")
+        async def get_voice_library_endpoint(request):
+            """Return current dropdown keys for Character Voices."""
+            try:
+                voice_discovery_module = get_voice_discovery_module()
+                force_refresh = request.query.get("refresh", "0").strip().lower() in {"1", "true", "yes"}
+                voices = voice_discovery_module.get_available_voices(force_refresh=force_refresh)
+                response = web.json_response({"voices": voices})
+                response.headers["Cache-Control"] = "no-store, no-cache, must-revalidate"
+                return response
+            except Exception as e:
+                print(f"⚠️ Error serving voice library: {e}")
+                return web.json_response({"error": str(e)}, status=500)
+
+        @PromptServer.instance.routes.get("/api/tts-audio-suite/voice-preview")
+        async def get_voice_preview_endpoint(request):
             """
             Stream selected Character Voices dropdown audio for browser preview playback.
 
@@ -473,15 +594,7 @@ print(json.dumps({"devices": devices}))
                 if not voice_name or voice_name == "none":
                     return web.json_response({"error": "voice_name is required and cannot be 'none'"}, status=400)
 
-                # Load voice discovery directly by file path to avoid package import issues
-                voice_discovery_path = os.path.join(os.path.dirname(__file__), "utils", "voice", "discovery.py")
-                spec = importlib.util.spec_from_file_location("voice_discovery_module", voice_discovery_path)
-                voice_discovery_module = importlib.util.module_from_spec(spec)
-                spec.loader.exec_module(voice_discovery_module)
-
-                # Use cached discovery for fast preview playback.
-                voice_discovery_module.get_available_voices(force_refresh=False)
-                audio_path, _ = voice_discovery_module.load_voice_reference(voice_name)
+                audio_path, _ = resolve_character_voice(voice_name)
 
                 if not audio_path or not os.path.exists(audio_path):
                     return web.json_response({"error": f"Voice file not found: {voice_name}"}, status=404)
@@ -491,8 +604,30 @@ print(json.dumps({"devices": devices}))
                 response.headers["Cache-Control"] = "no-store, no-cache, must-revalidate"
                 return response
             except Exception as e:
-                print(f"⚠️ Error serving voice preview audio: {e}")
-                return web.json_response({"error": str(e)}, status=500)
+                print(f"⚠️ Error serving voice preview audio: {e}")
+                return web.json_response({"error": str(e)}, status=500)
+
+        @PromptServer.instance.routes.get("/api/tts-audio-suite/voice-info")
+        async def get_voice_info_endpoint(request):
+            """Return canonical metadata for a Character Voices dropdown entry."""
+            try:
+                voice_name = request.query.get("voice_name", "").strip()
+                if not voice_name or voice_name == "none":
+                    return web.json_response({"error": "voice_name is required and cannot be 'none'"}, status=400)
+
+                audio_path, reference_text = resolve_character_voice(voice_name)
+                if not audio_path or not os.path.exists(audio_path):
+                    return web.json_response({"error": f"Voice file not found: {voice_name}"}, status=404)
+
+                response = web.json_response({
+                    "voice_name": voice_name,
+                    "reference_text": reference_text or "",
+                })
+                response.headers["Cache-Control"] = "no-store, no-cache, must-revalidate"
+                return response
+            except Exception as e:
+                print(f"⚠️ Error serving voice metadata: {e}")
+                return web.json_response({"error": str(e)}, status=500)
 
         @PromptServer.instance.routes.post("/api/tts-audio-suite/audio-analyzer-preview")
         async def audio_analyzer_preview_endpoint(request):

docs/ADVANCED_TTS_TAG_EDITOR_NODE.md

@@ -2,7 +2,7 @@
 
 ## Overview
 
-Advanced multiline string editor node that extends ComfyUI's standard multiline widget with a sidebar containing context-aware controls for TTS-specific tags (character switching, parameters, pauses). The node features real-time tag generation, preset management, and intelligent syntax support based on the CHARACTER_SWITCHING_GUIDE and PARAMETER_SWITCHING_GUIDE.
+Advanced multiline string editor node that extends ComfyUI's standard multiline widget with a sidebar containing context-aware controls for TTS-specific tags (character switching, parameters, pauses, and engine-native inline controls). The node features real-time tag generation, preset management, engine-aware quick swapping, and intelligent syntax support based on the CHARACTER_SWITCHING_GUIDE and PARAMETER_SWITCHING_GUIDE.
 
 ---
 
@@ -85,10 +85,14 @@ Advanced multiline string editor node that extends ComfyUI's standard multiline
 - Split by paragraph breaks or sentence punctuation
 - Batch apply parameters across multiple `[Character]text` blocks
 
-#### Tag Inspector
-- Show existing tags in selection/current line
-- Checkbox UI to toggle tags on/off temporarily
-- Quick-edit dialog for existing tag values
+#### Tag Inspector
+- Show existing tags in selection/current line
+- Checkbox UI to toggle tags on/off temporarily
+- Click a character, language, audio-reference, parameter, or supported engine-native inline tag to open a color-coded quick-swap palette
+- Click a palette option once to keep the palette open, then click again to commit; press-and-hold, drag, and release selects in one gesture
+- Palette choices follow the selected inline engine (including IndexTTS-2, Higgs Audio v3, Step Audio EditX, CosyVoice3, and OmniVoice)
+- Quoted IndexTTS-2 text emotion tags remain direct editable text and do not open a replacement palette
+- Long bracket and angle tags wrap inside the editor rather than overflowing horizontally
 
 #### Auto-Formatting
 - Button: "Auto-Format Tags" → organize tags consistently

docs/BUMP_SCRIPT_INSTRUCTIONS.md

@@ -12,21 +12,26 @@
 
 **⚠️ IMPORTANT: Use positional arguments, NOT --commit/--changelog flags**
 
-```bash
-# EASIEST: Just use 'patch' - script auto-increments the version
-python3 scripts/bump_version_enhanced.py patch "<commit_desc>" "<changelog_desc>"
+```powershell
+# Windows: use the canonical ComfyUI environment for this project
+& 'J:\stablediffusion1111s2\Data\Packages\ComfyUIPy129\test_env_error\Scripts\python.exe' 'scripts\bump_version_enhanced.py' patch "<commit_desc>" "<changelog_desc>"
 
-# OR: Specify exact version if needed
-python3 scripts/bump_version_enhanced.py <version> "<commit_desc>" "<changelog_desc>"
+# OR: specify an exact version if needed
+& 'J:\stablediffusion1111s2\Data\Packages\ComfyUIPy129\test_env_error\Scripts\python.exe' 'scripts\bump_version_enhanced.py' <version> "<commit_desc>" "<changelog_desc>"
+```
+
+```bash
+# Linux/macOS: use the available Python 3 interpreter
+python3 scripts/bump_version_enhanced.py patch "<commit_desc>" "<changelog_desc>"
 ```
 
 ### Examples
 
 #### Multiline Format (Recommended Standard)
 
-```bash
+```powershell
 # Patch release (bug fixes) - CORRECT FORMAT
-python3 scripts/bump_version_enhanced.py 3.2.9 "Fix character alias resolution
+& 'J:\stablediffusion1111s2\Data\Packages\ComfyUIPy129\test_env_error\Scripts\python.exe' 'scripts\bump_version_enhanced.py' 3.2.9 "Fix character alias resolution
 
 Technical details:
 - Fix parser bypassing character tags in single mode
@@ -37,8 +42,8 @@ Technical details:
 - Improve character name recognition accuracy
 - Better error handling for invalid character names"
 
-# Minor release (new features) - CORRECT FORMAT  
-python3 scripts/bump_version_enhanced.py 3.3.0 "Add Higgs Audio 2 TTS engine
+# Minor release (new features) - CORRECT FORMAT
+& 'J:\stablediffusion1111s2\Data\Packages\ComfyUIPy129\test_env_error\Scripts\python.exe' 'scripts\bump_version_enhanced.py' 3.3.0 "Add Higgs Audio 2 TTS engine
 
 Implementation details:
 - Integrate boson_multimodal voice cloning system
@@ -50,7 +55,7 @@ Implementation details:
 - Multiple built-in voice presets available"
 
 # Major release (breaking changes) - CORRECT FORMAT
-python3 scripts/bump_version_enhanced.py 4.0.0 "Complete unified architecture implementation
+& 'J:\stablediffusion1111s2\Data\Packages\ComfyUIPy129\test_env_error\Scripts\python.exe' 'scripts\bump_version_enhanced.py' 4.0.0 "Complete unified architecture implementation
 
 Breaking changes:
 - Migrate all nodes to unified interface pattern
@@ -64,22 +69,22 @@ Breaking changes:
 ```
 
 #### Auto-Increment Examples (Recommended)
-```bash
-# Auto-increment patch version (4.5.25 → 4.5.26) - CORRECT FORMAT
-python3 scripts/bump_version_enhanced.py patch "Fix character parsing issues" "Fix character name handling in TTS generation"
+```powershell
+# Auto-increment patch version (4.5.25 → 4.5.26) - Windows
+& 'J:\stablediffusion1111s2\Data\Packages\ComfyUIPy129\test_env_error\Scripts\python.exe' 'scripts\bump_version_enhanced.py' patch "Fix character parsing issues" "Fix character name handling in TTS generation"
 
-# Auto-increment minor version (4.5.25 → 4.6.0) - CORRECT FORMAT  
-python3 scripts/bump_version_enhanced.py minor "Add new TTS engine support" "Add Higgs Audio 2 TTS engine with voice cloning"
+# Auto-increment minor version (4.5.25 → 4.6.0) - Windows
+& 'J:\stablediffusion1111s2\Data\Packages\ComfyUIPy129\test_env_error\Scripts\python.exe' 'scripts\bump_version_enhanced.py' minor "Add new TTS engine support" "Add Higgs Audio 2 TTS engine with voice cloning"
 ```
 
 #### Single-Line Format (Only for Super Minor Changes)
-```bash
-python3 scripts/bump_version_enhanced.py patch "Fix typo in node tooltip" "Fix typo in audio analyzer tooltip"
+```powershell
+& 'J:\stablediffusion1111s2\Data\Packages\ComfyUIPy129\test_env_error\Scripts\python.exe' 'scripts\bump_version_enhanced.py' patch "Fix typo in node tooltip" "Fix typo in audio analyzer tooltip"
 ```
 
 #### Dry-Run Preview (Test Before Committing)
-```bash
-python3 scripts/bump_version_enhanced.py patch "Fix preview issues" "Fix preview not reflecting filter parameters" --dry-run
+```powershell
+& 'J:\stablediffusion1111s2\Data\Packages\ComfyUIPy129\test_env_error\Scripts\python.exe' 'scripts\bump_version_enhanced.py' patch "Fix preview issues" "Fix preview not reflecting filter parameters" --dry-run
 ```
 
 #### Auto-Categorization System
@@ -99,21 +104,22 @@ python3 scripts/bump_version_enhanced.py patch "Fix preview issues" "Fix preview
   - **Commit**: Technical implementation details for developers
   - **Changelog**: User-facing benefits and impacts
 
-**Bash Syntax Notes:**
+**Command Syntax Notes:**
 - Multiline strings need proper quoting (opening quote on first line, closing quote on last line)
-- Use `\` (backslash) for line continuation in bash commands
+- The Windows command uses PowerShell's `&` call operator and the canonical project Python path
+- The Linux/macOS command uses `python3`
 - Don't add manual category prefixes like "Fixed:" - script handles categorization automatically!
 
 ### Interactive Mode (Recommended for Complex Changes)
 
-```bash
-python3 scripts/bump_version_enhanced.py 3.2.9 --interactive
+```powershell
+& 'J:\stablediffusion1111s2\Data\Packages\ComfyUIPy129\test_env_error\Scripts\python.exe' 'scripts\bump_version_enhanced.py' 3.2.9 --interactive
 ```
 
 ### Legacy Mode (Same Description for Both)
 
-```bash
-python3 scripts/bump_version_enhanced.py 3.2.9 "Fix bugs and improve stability"
+```powershell
+& 'J:\stablediffusion1111s2\Data\Packages\ComfyUIPy129\test_env_error\Scripts\python.exe' 'scripts\bump_version_enhanced.py' 3.2.9 "Fix bugs and improve stability"
 ```
 
 ### What the Script Does
@@ -239,13 +245,14 @@ git commit -m "Prepare for version bump"
 - Use semantic versioning: `4.5.25` (not `v4.5.25` or `4.5`)
 - Or use auto-increment: `patch`, `minor`, `major`
 
-**Bash syntax errors with multiline**
-- Make sure opening quote is on same line as `--commit` or `--changelog`  
+**Command syntax errors with multiline**
+- Make sure opening quote is on same line as `--commit` or `--changelog`
 - Make sure closing quote is on its own line
-- Use `\` for line continuation
+- On Windows, use the PowerShell command shown above
+- On Linux/macOS, use `python3`
 
 **Want to see what will happen before committing?**
-```bash
+```powershell
 # Add --dry-run to preview changelog categorization
-python3 scripts/bump_version_enhanced.py patch "description" "changelog" --dry-run
-```
+& 'J:\stablediffusion1111s2\Data\Packages\ComfyUIPy129\test_env_error\Scripts\python.exe' 'scripts\bump_version_enhanced.py' patch "description" "changelog" --dry-run
+```

docs/COSYVOICE3_TAGS_GUIDE.md

@@ -4,7 +4,9 @@
 
 CosyVoice3 has **native paralinguistic tag support** built into the model. Unlike Step Audio EditX post-processing, these tags are processed **during TTS generation** for natural integration.
 
-**⚠️ IMPORTANT**: Step Audio EditX post-processing is **DISABLED** for CosyVoice3. EditX syntax like `<Laughter:2>`, `<emotion:happy>`, `<style:whisper>` will be stripped but **not processed**. Use CosyVoice3 native tags instead.
+**⚠️ IMPORTANT**: Step Audio EditX post-processing is **DISABLED** for CosyVoice3. EditX syntax like `<Laughter:2>`, `<emotion:happy>`, `<style:whisper>` will be stripped but **not processed**. Use CosyVoice3 native tags instead.
+
+The multiline TTS tag editor has an engine-aware `Inline Tags` panel with a dedicated `CosyVoice3` mode that inserts these native tags directly.
 
 ---
 
@@ -216,7 +218,7 @@ A: Yes! Include tags in subtitle text.
 
 ## Related Guides
 
-- **[Inline Edit Tags Guide](INLINE_EDIT_TAGS_USER_GUIDE.md)** - Step Audio EditX tags for other engines
+- **[Step Audio EditX Inline Tags Guide](INLINE_EDIT_TAGS_USER_GUIDE.md)** - Step Audio EditX tags for other engines
 - **[Parameter Switching Guide](PARAMETER_SWITCHING_GUIDE.md)** - Per-segment parameter control
 
 ---

docs/DRAMABOX_LORA_GUIDE.md

@@ -0,0 +1,99 @@
+# DramaBox LoRA training
+
+TTS Audio Suite exposes the official DramaBox audio-branch IC-LoRA trainer
+through the unified `🎓 Model Training` flow. The bundled scripts are pinned to
+the same upstream DramaBox revision as the inference implementation.
+
+See the official DramaBox
+[LoRA training guide](https://github.com/resemble-ai/DramaBox#training-a-lora-on-top-of-dramabox)
+for the upstream dataset format and training behavior.
+
+## Workflow
+
+1. Build a `⚙️ DramaBox Engine`.
+2. Create the dataset either externally or entirely inside ComfyUI:
+   `🎞️ Training Clip Staging` → `🧾 DramaBox Dataset Rows`.
+3. Connect the resulting manifest to `📦 DramaBox Dataset Prep` and keep
+   `dataset_type` set to `manifest`.
+4. Provide at least two clips per speaker.
+5. Connect the dataset to `🎛️ DramaBox Training Config` and then to `🎓 Model Training`.
+6. Select the resulting adapter in the DramaBox engine, or enter its path in the
+   advanced LoRA override field.
+
+The dataset node accepts:
+
+- JSONL/JSON manifests with `audio_filepath` (or `audio_path`) and `text` (or
+  `transcript`)
+- TSV rows with audio path and text
+- the official `gemini_synthetic` and `libriheavy` index formats
+
+Manifest rows may include `speaker`, `speaker_id`, `language`, and `duration`.
+If `speaker` is omitted, rows are grouped as `speaker_1`. Duration and audio
+metadata are measured without loading the waveform into the GPU. The suite
+converts all accepted formats into the `~`-delimited speaker index required by
+the upstream training loop. Clips are restricted to 2–20 seconds by default.
+
+For an all-ComfyUI dataset, connect one or more `AUDIO` sources to
+`🎞️ Training Clip Staging`, then enter one transcript per clip in
+`🧾 DramaBox Dataset Rows`. Speaker and language lines are optional; shared
+defaults are used when those lines are blank.
+
+### Transcripts and scene descriptions
+
+The official trainer accepts either plain spoken transcripts or the same
+scene-style prompt format used for inference. For example, both of these are
+valid training text:
+
+```text
+This is the spoken sentence.
+A woman speaks warmly, "This is the spoken sentence."
+```
+
+Use scene descriptions only when they accurately describe the clip. Plain
+transcripts remain valid and are the safer choice when no reliable style or
+scene annotation is available.
+
+## What training does
+
+The first preprocessing pass uses Gemma and the DramaBox audio VAE to create
+cached conditions and audio latents. The training process then attaches a LoRA
+to the audio transformer branch. It saves periodic checkpoints and exports the
+selected adapter to:
+
+```text
+ComfyUI/models/TTS/dramabox/loras/<adapter_name>/
+```
+
+The job directory, normalized index, preprocessing cache, progress file, and
+logs are stored under:
+
+```text
+ComfyUI/output/tts_audio_suite_training/dramabox/
+```
+
+`continue_from` is a warm start from an existing LoRA checkpoint; it is not an
+exact optimizer-state resume. Use saved checkpoints to compare quality rather
+than assuming the last step is best. Optional upstream validation can be
+enabled with a `val_config` YAML path, but it launches full DramaBox inference
+at each save step. It requires a second GPU: set `validation_gpu` to that
+physical CUDA device index. The suite rejects validation on the training GPU
+instead of allowing both full model processes to compete for the same VRAM.
+
+DramaBox LoRA inference supports normal transformer precision, `fp8_cast`, and
+the optional `torch.compile` path. With normal precision the live adapter is
+reversibly merged for fast inference. With FP8 storage the BF16 adapter remains
+unmerged above the immutable FP8 base weights, avoiding unsafe mixed-dtype
+weight fusion while retaining the main FP8 memory saving.
+
+The base DramaBox runtime is reused when the selected adapter or LoRA strength
+changes. Strength updates are applied directly to the live PEFT adapter, while
+the generated-audio cache still treats adapter path, file revision, and strength
+as distinct generation settings. Replacing an adapter with a different rank may
+retrace compiled transformer blocks, but does not reload the base checkpoint.
+
+## CPU-safe preflight
+
+Training and Gemma/VAE preprocessing are GPU workloads. For development or
+validation without touching CUDA, enable `dry_run` in the training config and
+the dataset node's `dry_run`/`preprocess_now` controls. This writes the
+normalized index and official command/config without loading DramaBox weights.

docs/DRAMABOX_PROMPTING_GUIDE.md

@@ -0,0 +1,163 @@
+# DramaBox Prompting Guide
+
+DramaBox is an English expressive TTS engine. It accepts ordinary narration,
+dialogue in quotation marks, and natural-language stage directions in one
+prompt.
+
+## Basic prompts
+
+Use quoted text for speech and surrounding prose for delivery:
+
+```text
+A tired detective speaks quietly in a rain-soaked office. "I knew this case would find me again."
+```
+
+`prompt_template` defaults to `"{seg}"`. `{seg}` is replaced by the current
+plain fragment, so the default marks the whole fragment as literal spoken
+dialogue. For example, `Hello.` becomes `"Hello."`. Clear the template field to
+send plain text unchanged.
+
+Customize the template to add delivery context, for example
+`A man speaks warmly, "{seg}"`. Quote-only input is normalized without adding
+another pair of quotes. Complete scene prompts with directions outside their
+quotation marks remain unchanged. Every non-empty template must contain `{seg}`.
+If it is omitted accidentally, TTS Audio Suite warns once and appends
+`"{seg}"` automatically instead of failing generation.
+
+For a one-segment override, `prompt_template` (or its `template` alias)
+automatically enables templating for that segment and then reverts to the node
+setting:
+
+```text
+[Narrator|template:A woman whispers, "{seg}"] This line uses a custom wrapper.
+```
+
+DramaBox can render non-verbal and delivery cues when they are described
+naturally:
+
+```text
+She tries to stay serious, then breaks into a short laugh. "That is the worst excuse I have ever heard." She sighs and continues more gently. "But I believe you."
+```
+
+Write one continuous scene paragraph. DramaBox does not require newlines as
+prompt syntax. In TTS Audio Suite, an untagged newline starts another generated
+segment, so use prose action directions and quoted dialogue in the same
+paragraph when they should remain one coherent DramaBox scene.
+
+Do not use ChatterBox V2 special tokens such as `[giggle]`. DramaBox was
+trained for prose-style scene direction, not that token vocabulary.
+
+## Voice references
+
+Reference audio is optional. Connect narrator audio or use a character voice
+file to clone its speaker and delivery. Upstream uses the first 10 seconds, so
+a clean single-speaker clip is the useful input; a transcript is not required.
+
+Without a reference, DramaBox uses its built-in voice behavior.
+
+DramaBox can occasionally produce a near-silent sample for a particular
+combination of reference audio, reference duration, generation duration, and
+seed. TTS Audio Suite checks the decoded waveform and prints a warning when both
+its RMS and peak levels are conservatively near silence. The audio is preserved;
+the suite does not retry or change parameters automatically. Try another
+generation duration, reference duration/audio, guidance setting, or seed for
+the affected segment. A different seed can help some combinations but is not a
+guaranteed fix.
+
+The warning is also propagated to node outputs. TTS Text includes affected
+segments in `generation_info`. TTS SRT marks affected subtitle numbers in
+`timing_report`, including the parameters that may be worth testing for that
+segment.
+
+## Character and pause tags
+
+TTS Audio Suite character tags still work. Each tagged character is generated
+as a separate DramaBox segment:
+
+```text
+[Alice] "We should leave now."
+[Bob] He answers without looking up. "Give me one minute."
+[pause:0.8]
+[Alice] "You said that five minutes ago."
+```
+
+Suite pause tags create exact silence outside the model. Natural pauses inside
+a spoken scene are better expressed in the prose prompt.
+
+## Engine controls
+
+- `cfg_scale`: text/prompt guidance. Official default: `2.5`.
+- `stg_scale`: skip-token guidance. Official default: `1.5`.
+- `duration_multiplier`: scales the estimated speaking duration. Official
+  default: `1.1`.
+- `gen_duration`: explicit generated-audio duration from `0` to `60` seconds.
+  `0` keeps automatic prompt-based estimation.
+- `ref_duration`: uses the first `3` to `30` seconds of a voice reference.
+  The default is `10`; audio later in the source file is ignored.
+- `rescale_scale`: CFG latent rescaling. Use `auto` or a fixed value from
+  `0` to `1`.
+- `watermark`: enables the optional official Perth output watermark. It is off
+  by default and requires Perth.
+- `seed`: supplied by the unified TTS Text or SRT node.
+
+Segment overrides support `seed`, `cfg_scale`, `stg_scale`, and
+`duration_multiplier`, `gen_duration`, `ref_duration`, and `rescale_scale`.
+Watermarking remains a whole-engine setting rather than a segment override.
+
+DramaBox performs its own duration-aware long-form chunking. The suite does
+not split a DramaBox scene by character count before passing it to the model.
+Automatically estimated scenes above 45 seconds use text chunking. A nonzero
+`gen_duration` remains one native generation so its explicit 0–60 second
+target is preserved.
+
+The unified SRT node's **Native Duration Targeting** option passes each
+subtitle's duration to DramaBox before final timing assembly. For subtitles
+containing multiple character or pause-separated fragments, the available
+speech time is allocated proportionally after explicit pause durations and
+inline `gen_duration` overrides are accounted for. The selected SRT timing
+mode still performs its normal final correction.
+
+## Negative Prompt and Segment Switching
+
+DramaBox uses CFG and exposes its negative prompt in the engine node. The
+default discourages robotic, distorted, noisy, muffled, unclear, and monotone
+speech. Override it for one character segment with:
+
+```text
+[Alice|negative:robotic, muffled] "Keep this line clean and intimate."
+[Bob|neg:noise, static] "This line uses a different negative prompt."
+```
+
+The segment override ends at the next character tag.
+
+## Memory and Performance
+
+- `fast` keeps all components on CUDA for the fastest repeated generation.
+- `staged` is an experimental strategy for lowering peak VRAM. It loads and
+  releases Gemma, the voice encoder, and audio decoder by stage, at the cost of
+  reloading them for each generated segment or long-form chunk.
+- `sequential` is a more aggressive experimental strategy for lowering peak
+  VRAM. It additionally keeps the diffusion transformer in system RAM while
+  another major stage uses CUDA. It transfers the transformer for every
+  generated segment or long-form chunk and is therefore substantially slower.
+  Actual peak usage varies with the environment, generation settings, and
+  other loaded components; no minimum GPU size is guaranteed.
+  System RAM must hold the offloaded transformer (about 3.4GB with FP8 or
+  6.6GB without it).
+- `fp8_cast` uses the official LTX FP8 transformer weight-storage policy and
+  upcasts linear weights during inference. It can lower VRAM and may be slower.
+- `compile_model` compiles the diffusion transformer blocks with DramaBox's
+  bundled LTX compilation path. The first generation can take substantially
+  longer while kernels compile; later denoising may be faster.
+
+## Requirements and license
+
+The full download is approximately 16.4GB and the official runtime requires
+an NVIDIA CUDA GPU. Fast mode targets roughly 24GB VRAM; the experimental
+staged modes can run with less memory at a speed cost. Output is 48kHz stereo.
+The optional official Perth watermark is applied only when enabled and the
+dependency is available.
+
+DramaBox uses the LTX-2 Community License. Entities with at least USD 10
+million in annual revenue require a separate paid commercial license. Review
+the bundled license before production use.

docs/Dev reports/DRAMABOX_CHATTERBOX_V3_CAPABILITY_SCOPE.md

@@ -0,0 +1,183 @@
+# DramaBox and Chatterbox Multilingual V3 Capability and Scope
+
+Research date: 2026-07-25
+
+## Official references
+
+- DramaBox code: `resemble-ai/DramaBox` at
+  `a70a5818e103c1c9fef22409c1e0c707ebf4f8a7`
+- DramaBox weights: `ResembleAI/Dramabox` at
+  `404f967f653fa1170dc15a9d1ddd3fdb9a0a842d`
+- Chatterbox code: `resemble-ai/chatterbox` at
+  `5de7a54aa4e5e2baadb0182dde554908b48b85c2`
+- Chatterbox weights: `ResembleAI/chatterbox` at
+  `5bb1f6ee58e50c3b8d408bc82a6d3740c2db6e18`
+- ComfyUI reference only: `kat3ri/ComfyUI-DramaBox` at
+  `715fcb11cc14d8c185438e2319b52fc00163941c`
+
+The repositories were cloned under
+`IgnoredForGitHubDocs/For_reference/`.
+
+## DramaBox capability report
+
+### Native scope
+
+- Task: English text-to-speech with optional zero-shot voice cloning.
+- Expressive control: prompt-driven speaker description, delivery, emotion,
+  pauses, laughs, sighs, and transitions.
+- Voice input: optional reference audio; upstream uses up to 10 seconds.
+- No native voice conversion, ASR, or audio editing API.
+- No language control. The official model is English-only.
+- No extra special node is required. Its structured scene prompt fits the
+  existing unified text and SRT nodes.
+
+### Native generation parameters
+
+- `cfg_scale` (official warm-server default `2.5`)
+- `stg_scale` (default `1.5`)
+- `duration_multiplier` (default `1.1`)
+- `seed` (default `42`)
+- `ref_duration` (default `10.0` seconds)
+- `rescale_scale` (`auto` by default)
+- `gen_duration` (`0` means automatic)
+- Official long-form chunk limits and crossfade parameters
+
+The initial suite UI exposes `cfg_scale`, `stg_scale`, and
+`duration_multiplier`. Seed remains owned by the unified TTS nodes.
+Reference duration, rescale, steps, modality guidance, and explicit output
+duration stay on official defaults because exposing them would add expert
+controls without a demonstrated suite use case. The official duration-aware
+long-form path is used automatically instead of adding duplicate chunk UI.
+
+### Audio and generation behavior
+
+- The LTX audio decoder returns stereo audio at 48 kHz.
+- The base model was trained on clips around 20 seconds. Current upstream
+  supports longer clips with a silence-prior correction and automatically
+  chunks prompts targeting about 37 seconds with a 45-second cap.
+- The official long-form chunker preserves the scene/speaker prefix and quote
+  groups, then joins chunks with a 50 ms equal-power crossfade.
+- Upstream applies the Perth watermark only in `generate_to_file()`, not in
+  the in-memory `generate()` method. The suite wrapper must therefore apply
+  the watermark to in-memory output explicitly.
+
+### Model layout
+
+Organized destination: `ComfyUI/models/TTS/dramabox/DramaBox/`
+
+- `dramabox-dit-v1.safetensors` — 6,575,225,528 bytes
+- `dramabox-audio-components.safetensors` — 1,942,831,020 bytes
+- `assets/silence_latent_frame.pt` — 1,501 bytes
+- `gemma-3-12b-it-bnb-4bit/`
+  - two safetensor shards plus tokenizer/config files from
+    `unsloth/gemma-3-12b-it-bnb-4bit`
+
+The implementation must use the suite downloader with `local_dir`-style
+organized downloads and disable Transformers/Hugging Face fallback downloads.
+
+### Dependencies and runtime
+
+The official requirements include Torch/Torchaudio 2.8, Transformers 4.45+,
+bitsandbytes 0.45+, Accelerate, PEFT, PyAV, Einops, SentencePiece,
+Safetensors, PyYAML, and Perth. The official source imports successfully in
+the configured suite validation environment with Torch 2.10 and Transformers
+5.10, so DramaBox belongs in the main Transformers 5 environment.
+
+The optional NVIDIA RE-USE reference denoiser is intentionally excluded:
+its Mamba dependencies have no practical Windows installation path and its
+NSCLv1 non-commercial license is a poor default for the suite.
+
+### License
+
+DramaBox code and weights are under the LTX-2 Community License, not MIT.
+The license requires attribution, use restrictions, modified-file notices,
+and a separate paid license for entities with at least USD 10 million in
+annual revenue. The upstream license must ship beside any bundled inference
+code, and the engine UI/docs must disclose the restriction.
+
+## Existing ComfyUI reference notes
+
+`kat3ri/ComfyUI-DramaBox` confirms useful ComfyUI audio-shape handling,
+organized model paths, the warm `TTSServer` API, and practical UI ranges.
+It must not be copied as architecture:
+
+- It auto-clones source code at runtime.
+- It has no unified model lifecycle, cache, character/pause integration, SRT
+  processor, interrupt handling, or generation report integration.
+- It directly calls the engine from a standalone node.
+- It patches partially imported bitsandbytes modules globally.
+- Its README says output is watermarked, but its node calls the unwatermarked
+  in-memory upstream path.
+
+## Chatterbox Multilingual V3 capability report
+
+V3 is not a new engine. Official upstream loads it as an opt-in checkpoint through
+`ChatterboxMultilingualTTS.from_pretrained(..., t3_model="v3")`; the only
+model-family change is selecting `t3_mtl23ls_v3.safetensors` instead of the
+V2 T3 checkpoint. Its official generation path also skips the legacy
+alignment analyzer, uses repetition penalty `1.2`, and removes the final
+degraded pre-EOS speech-token artifact. It keeps
+the same 500M architecture, 23-language list, 24 kHz output, voice-reference
+mode, tokenizer, voice encoder, S3Gen decoder, and generation parameters:
+
+- `language_id`
+- `exaggeration`
+- `cfg_weight`
+- `temperature`
+- `repetition_penalty`
+- `min_p`
+- `top_p`
+
+The suite forwards V3 `exaggeration` using the upstream/native scale. Manual
+testing found little or no audible response across values, so this remains a
+current checkpoint limitation rather than a suite-side scaling issue.
+
+The existing `chatterbox_official_23lang` engine already implements Unified
+TTS Text, Unified SRT TTS, voice references, caching, character switching,
+pause tags, parameter switching, and lifecycle handling. V3 therefore
+extends its `model_version` choices and downloader requirements. It must not
+create a second engine node or duplicate processors.
+
+## Integration scope
+
+### DramaBox
+
+- Unified TTS Text: yes
+- Unified SRT TTS: yes
+- Character tags and narrator fallback: yes
+- Pause tags: yes
+- Segment switching: `seed`, `cfg_scale`, `stg_scale`, and
+  `duration_multiplier`
+- Generated audio cache: yes
+- Long-form strategy: official duration-aware chunker; ignore suite
+  character-count chunking inside each already separated character/pause
+  segment
+- Clear VRAM: full TTSServer teardown and lazy reload because the quantized
+  Gemma stack should not be copied to system RAM
+- Runtime: main environment
+- Voice Changer / ASR / editing / special node: no
+
+### Chatterbox Multilingual V3
+
+- Extend existing Official 23-Lang model version control with V3.
+- Keep V2 available for backward compatibility.
+- Make V3 the suite default for new configurations so the newly requested
+  version is immediately selected. Upstream still defaults to V2 and exposes
+  V3 as opt-in.
+- Existing saved workflows with V1/V2 values continue to load unchanged.
+
+## Validation matrix
+
+- Static import and registration checks
+- Chatterbox V1/V2/V3 file-resolution tests without downloading weights
+- DramaBox downloader layout checks without downloading the 16+ GB models
+- DramaBox TTS processor tests with a fake adapter for character, pause,
+  cache-facing parameter, audio-shape, and combination behavior
+- SRT processor interrupt and timing-path tests with fake generation
+- Live FL-MCP checks after restarting ComfyUI:
+  - engine and unified nodes register
+  - smallest DramaBox text workflow loads
+  - smallest DramaBox SRT workflow loads
+  - full generation is attempted only if all 16+ GB weights and adequate
+    VRAM are available
+- Human assessment remains required for subjective audio quality.

docs/Dev reports/ISOLATED_RUNTIMES_PLAN.md

@@ -0,0 +1,67 @@
+# Isolated Runtimes Plan
+
+## Goal
+
+Keep ComfyUI in one modern main runtime while routing fragile engines into
+their own pinned Python environments.
+
+## Why
+
+- `transformers` major-version breaks now hit engines differently
+- one broken HF stack should not poison the whole node runtime
+- per-engine rollback is simpler than global dependency rollback
+
+## User impact
+
+- isolated runtimes duplicate dependencies on disk
+- large packages such as PyTorch will usually be installed again inside the worker runtime
+- this must be warned in:
+  - docs
+  - node tooltips
+  - runtime creation logs
+
+## Contract
+
+ComfyUI process:
+- owns UI, node orchestration, model discovery, caching policy
+- prepares normalized request payloads
+- launches a worker only for engines configured as `runtime_mode=isolated`
+
+Worker process:
+- runs in a dedicated Python env
+- receives a JSON request
+- performs load / generate / unload actions
+- returns JSON response + output artifact paths
+
+## Initial profiles
+
+- `vibevoice_transformers4`
+- `qwen3_tts_transformers5`
+- `step_audio_editx_transformers5`
+- `moss_tts_transformers5`
+
+## First target
+
+VibeVoice / KugelAudio
+
+Reason:
+- highest dependency fragility
+- strongest evidence that global `transformers 5` compatibility is not enough
+- likely template for future HF-heavy engines
+
+## Implementation order
+
+1. Add config + registry support for runtime routing
+2. Add launcher + request/response protocol
+3. Add VibeVoice worker entrypoint
+4. Add opt-in routing from unified interface
+5. Add user-facing runtime path configuration
+6. Expand to Qwen / MOSS / Step Audio only after VibeVoice works
+
+## Non-goals
+
+- no silent env creation
+- no arbitrary command execution
+- no shell-constructed user input
+- no automatic migration of all engines
+- no pretending disk duplication does not exist

docs/Dev reports/TODO.md

@@ -2,10 +2,17 @@
 
 This document tracks architectural issues and inconsistencies that need refactoring to improve modularity and reduce code duplication.
 
-## SRT Processing Architecture Issues
-
-### Problem: Inconsistent SRT Implementation Approaches
-Different engines use completely different patterns for SRT processing:
+## SRT Processing Architecture Issues
+
+### Problem: VibeVoice Native Multi-Speaker SRT Uses Non-Contiguous Global Slots
+- **Issue**: Later subtitles can request global `Speaker 2`/`Speaker 3` IDs while VibeVoice truncates and renumbers the supplied voice prompts from zero.
+- **Impact**: A subtitle containing only later global speakers can bind the wrong reference or leave a requested speaker without a matching voice prompt.
+- **Solution**: Use the global character map only to select references, then compact each subtitle to request-local speakers `0..N-1` with references in the same order.
+- **Regression test**: Cover subtitle 1 `[Alice]...[Bob]...`, followed by subtitle 2 `[Bob]...[Rick]...`.
+- **Priority**: High
+
+### Problem: Inconsistent SRT Implementation Approaches
+Different engines use completely different patterns for SRT processing:
 
 1. **ChatterBox (Old)**: Uses `ChatterboxSRTTTSNode` class (should be processor)
 2. **VibeVoice**: Uses proper `VibeVoiceSRTProcessor` class with full implementation
@@ -136,4 +143,4 @@ Each engine reimplements:
 ---
 
 *Last Updated: 2025-01-XX*
-*Add new issues to this file as they are discovered during development*
+*Add new issues to this file as they are discovered during development*

docs/Dev reports/TRANSFORMERS_5_HIGGS_AUDIO_REPORT.md

@@ -0,0 +1,123 @@
+TTS Audio Suite - Transformers 5.x Higgs Audio 2 Investigation Report
+
+Scope
+- Problem: Higgs Audio 2 regressed after the suite moved the main environment to Transformers 5.x.
+- Goal: determine whether Higgs Audio 2 is a good native Transformers 5 patch target, or whether it should move to runtime isolation.
+
+Environment
+- Windows / ComfyUI runtime.
+- Main env on Transformers 5.x.
+- Direct repro and validation performed in the real Windows Python environment outside the ComfyUI UI wrapper.
+
+Findings Summary
+- Higgs Audio 2 is not blocked at the shallow import/API level only.
+- Some real compatibility fixes were possible:
+  - bundled cache API adaptation
+  - cache-arg rename fixes
+  - major cold-load regression fix
+- However, the remaining failure is decode-path semantics, especially around StaticCache and CUDA graph replay.
+- Higgs can be made to load and "run" on Transformers 5, but the high-performance path is not trustworthy and produced incorrect audio.
+- This is not a good short-path patch target like IndexTTS.
+
+Reproduction
+1. Transformers 5 main environment.
+2. Higgs Audio 2 loaded through the normal TTS Text path with voice cloning.
+3. Observed one or more of:
+   - multi-minute cold-load stalls
+   - StaticCache API crashes
+   - CUDA graph capture/runtime failures
+   - runaway/gibberish generation
+   - near-empty or silent output
+
+Key Diagnostic Evidence
+- Initial cold-load timing showed:
+  - `⏱️ Higgs Audio: Core model load took 440.61s`
+- Tokenizer/HuBERT stages were not the bottleneck:
+  - text tokenizer around 1-2s
+  - audio tokenizer around 2-5s
+- The 440s stall was inside `HiggsAudioModel.from_pretrained(...)`, specifically Hugging Face missing-key initialization.
+- Direct checkpoint/model key comparison in runtime showed:
+  - `missing_count 0`
+  - `unexpected_count 0`
+- After bypassing unnecessary missing-key initialization for complete local checkpoints, direct timing dropped to:
+  - `⏱️ Higgs Audio: Core model load took 58.46s`
+  - later direct runs around `33-58s` depending on cache state
+
+- StaticCache/API drift evidence:
+  - `'StaticCache' object is not subscriptable`
+  - later mask/cache shape mismatches around static cache width `1024`
+
+- Cache-arg drift evidence:
+  - bundled Higgs/LLaMA path mixed `past_key_value` and `past_key_values`
+  - led to duplicated kwarg failures and broken decode semantics
+
+- CUDA graph replay evidence:
+  - direct runtime probe with graph replay active produced:
+    - immediate or near-immediate audio termination
+    - tiny garbage clips
+    - or invalid decode behavior
+  - direct runtime probe with StaticCache but without CUDA graph replay produced:
+    - normal-length audio token generation
+    - reasonable token spread
+    - real waveform output
+
+Attempts and Results (chronological)
+1) StaticCache compatibility shims
+   - Added bundled helpers to access cache tensors/device across old tuple caches and new Cache/StaticCache APIs.
+   - Result: fixed immediate StaticCache indexing crashes.
+
+2) CUDA graph failure fallback cleanup
+   - On graph capture failure, forced cache recreation and switched to DynamicCache instead of continuing with stale StaticCache assumptions.
+   - Result: removed some follow-on crashes after graph failure, but not the core correctness issue.
+
+3) Attention backend adjustment
+   - Stopped forcing eager attention on CUDA and preferred SDPA instead.
+   - Result: better alignment with the intended high-performance path, but did not fully restore correctness.
+
+4) TTS defaults forcing audio generation
+   - Set Higgs TTS defaults to force audio generation rather than letting the model drift in text mode.
+   - Result: prevented obvious free-running text-style generation, but did not solve bad output.
+
+5) Cold-load regression fix
+   - Added a local complete-checkpoint detector and skipped Hugging Face missing-key initialization when the local Higgs checkpoint was complete.
+   - Result: large real improvement to cold-load time.
+
+6) Cache-argument adaptation
+   - Patched bundled Higgs/LLaMA call sites for Transformers 5 cache-arg naming.
+   - Result: moved runtime further, but surfaced deeper semantic issues instead of fixing them outright.
+
+7) CUDA graph runner PyTorch compatibility
+   - Wrapped input-buffer updates in inference mode to satisfy current PyTorch restrictions.
+   - Result: removed one explicit runtime crash, but not the underlying decode corruption.
+
+8) StaticCache without CUDA graph replay direct comparison
+   - Direct Windows probes showed:
+     - `enable_cuda_graphs=True` with graph replay: corrupted/too-short audio output
+     - `enable_cuda_graphs=False`: long audio-token generation and real waveform output
+     - `StaticCache` with graph replay skipped: closer to sane behavior than replay path
+   - Result: isolated the remaining failure to the CUDA graph replay path on Transformers 5.
+
+Interpretation
+- Higgs Audio 2 on Transformers 5 is not blocked by a single missing symbol or a few renamed APIs.
+- The remaining problem is decode/runtime behavior, especially the interaction between:
+  - StaticCache
+  - bundled Higgs decode logic
+  - LLaMA attention/cache semantics on current Transformers 5
+  - CUDA graph replay
+- That is behavior-level drift, not just compatibility-shim work.
+
+Decision
+- Higgs Audio 2 on Transformers 5 is still theoretically patchable.
+- It is not a good use of time right now.
+- The engine reached the same bad zone as Qwen3-TTS, but with even more sensitivity around CUDA graphs and cache behavior.
+- Runtime isolation is the correct path for now.
+
+Operational Recommendation
+- Do not continue native Transformers 5 Higgs patching in the main branch.
+- Route Higgs Audio 2 into an isolated legacy runtime instead.
+- Prefer a dedicated Higgs runtime first rather than assuming shared legacy parity.
+- Revisit native Transformers 5 support only if Higgs becomes important enough to justify deeper decode/cuda-graph compatibility work.
+
+Notes
+- The cold-load fix was a real improvement and may still be worth revisiting later, even outside the final T5 decision.
+- The rest of this investigation should be treated as a branch-local investigation record, not as a stable main-branch solution.

docs/Dev reports/TRANSFORMERS_5_QWEN3_TTS_REPORT.md

@@ -1,4 +1,4 @@
-TTS Audio Suite - Transformers 5.0.0 Qwen3-TTS Investigation Report
+TTS Audio Suite - Transformers 5.x Qwen3-TTS Investigation Report
 
 Scope
 - Problem: Qwen3-TTS garbled audio or load failure on Transformers 5.0.0.
@@ -54,3 +54,65 @@ Recommendation
 
 Notes
 - No changes in this report alter project behavior.
+
+Addendum - Transformers 5.10.2 Re-check
+
+Scope
+- Problem: determine whether newer Transformers 5.10.2 improves Qwen3-TTS enough to avoid runtime isolation.
+- Goal: distinguish shallow API drift from deeper generation/runtime incompatibility.
+
+Environment
+- Windows / ComfyUI runtime.
+- Main env on Transformers 5.10.2.
+- Direct repro also performed in the real Windows Python environment outside the ComfyUI UI wrapper.
+
+What Changed vs 5.0.0
+- The failure mode is different.
+- The first blocker is no longer the tokenizer collapse described above.
+- Qwen3-TTS can be pushed significantly further on 5.10.2 with local compatibility shims.
+
+Key Diagnostic Evidence
+- Direct model load initially failed with:
+  - `KeyError: 'default'`
+  - source: `ROPE_INIT_FUNCTIONS[self.rope_type]`
+- Current Transformers 5.10.2 no longer exposes a legacy `"default"` RoPE entry:
+  - available keys observed in runtime: `dynamic`, `linear`, `llama3`, `longrope`, `proportional`, `yarn`
+- After restoring a local `"default"` compatibility entry, model load succeeded.
+- Generation then failed on changed Hugging Face masking APIs:
+  - `create_causal_mask() got an unexpected keyword argument 'input_embeds'`
+- After adapting those call sites to the current HF signature, generation progressed further.
+- Final direct generation failure on 5.10.2 was:
+  - `RuntimeError: Expected size for first two dimensions of batch2 tensor to be: [16, 277] but got: [16, 139].`
+  - source: SDPA attention path during generation.
+
+Attempts and Results (5.10.2)
+1) Legacy RoPE compatibility restore
+   - Added a local fallback for legacy `"default"` RoPE initialization.
+   - Result: Qwen3-TTS load progressed and completed successfully.
+2) Mask helper API adaptation
+   - Updated bundled Qwen mask call sites to match current HF `create_causal_mask` / `create_sliding_window_causal_mask` signatures.
+   - Result: generation progressed beyond the earlier immediate crash.
+3) Direct generation validation with real voice-clone inputs
+   - Used real reference audio and transcript in the Windows environment.
+   - Result: no longer a trivial import/load/signature issue; generation hit a tensor-shape mismatch inside SDPA.
+
+Interpretation
+- Transformers 5.10.2 is not blocked in the same way as 5.0.0.
+- Some problems are still shallow compatibility issues and can be adapted cleanly.
+- However, once generation reaches SDPA tensor-shape mismatch territory, this stops being a simple API-shim problem.
+- At that point, continuing the patch path means re-validating model behavior against newer Hugging Face generation/mask/cache semantics, not just restoring removed symbols.
+
+Decision
+- Qwen3-TTS on Transformers 5.10.2 is still technically patchable in principle.
+- It is not a good short-path target like IndexTTS.
+- The suite should prefer runtime isolation for Qwen3-TTS instead of continuing deep Transformers 5 surgery.
+
+Operational Recommendation
+- Treat Qwen3-TTS as a legacy-stack engine for now.
+- Route it into the shared legacy Transformers 4 runtime used by compatible engines (same family already proven for VibeVoice/Qwen in prior testing).
+- Keep the main environment on Transformers 5.
+- Revisit native Transformers 5 support only if Qwen3-TTS becomes strategically important enough to justify deeper behavior-level compatibility work.
+
+Notes
+- The transient 5.10.2 patch attempts were not kept as the main solution path.
+- The current project direction is runtime isolation, not continued in-place Transformers 5 patching for Qwen3-TTS.

docs/Dev reports/tts_audio_suite_aux_models.yaml

@@ -55,7 +55,6 @@ categories:
         languages:
           - "47 languages"
         notes: "Heavier multilingual fallback with the widest language coverage supported by the punctuators package."
-
 model_layouts_markdown: |
   # Auxiliary Model Layouts
 

docs/ENGINE_COMPARISON.md

@@ -2,18 +2,26 @@
 
 ## Engine Comparison
 
-| Engine             | Models                                    | Size         | TTS | SRT | VC  | ASR | Training | License                  | Special Features                                                                         | Languages                                                                                |
-| ------------------ | ----------------------------------------- | ------------ | :-: | :-: | :-: | :-: | :------: | ------------------------ | ---------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------- |
-| **F5-TTS**         | Base, v1, E2TTS + 8 lang models           | ~1.2GB each  | ✅ | ✅ | ❌ | ❌ | ❌ | CC-BY-NC-4.0             | Targeted Word/Speech Editing, Speed control                                              | 🇺🇸​🇩🇪​🇪🇸​🇫🇷​🇮🇹​🇯🇵​🇧🇷​🇵🇱​🇮🇳​🇹🇭                                                            |
-| **ChatterBox**     | EN, DE×3, IT, FR, RU, HY, KA, JA, KO, NO  | ~4.3GB       | ✅ | ✅ | ✅ | ❌ | ❌ | MIT                      | Expressiveness slider                                                                    | 🇺🇸​🇩🇪​🇫🇷​🇮🇹​🇯🇵​🇰🇷​🇷🇺​🇳🇴​🇦🇲​🇬🇪                                                            |
-| **ChatterBox 23L** | v1, v2, Vietnamese (Viterbox), Egyptian Arabic (oddadmix) | ~4.3GB       | ✅ | ✅ | ✅ | ❌ | ❌ | MIT                      | 24 languages in single model, emotion tokens (v2 - doesn't work)                         | 🇺🇸​🇨🇳​🇩🇪​🇪🇸​🇫🇷​🇮🇹​🇯🇵​🇰🇷​🇷🇺​🇵🇹​🇵🇱​🇮🇳​🇪🇬​🇹🇷​🇹🇭​🇳🇴​🇻🇳​🇩🇰​🇫🇮​🇬🇷​🇮🇱​🇲🇾​🇳🇱​🇸🇪​🇰🇪(+9)           |
-| **VibeVoice**      | 1.5B, 7B, KugelAudio-0 (7B), kugel-2 (7B), Hindi-1.5B/7B | 5.4GB / 18GB | ✅ | ✅ | ❌ | ❌ | ❌ | MIT (research-only per model card) | 90-min long-form, Native 4-speaker (Base models), Multilingual (KugelAudio variants), 4-bit quantization | 🇺🇸​🇨🇳​🇩🇪​🇪🇸​🇫🇷​🇮🇹​🇯🇵​🇰🇷​🇷🇺​🇧🇷​🇵🇱​🇮🇳​🇪🇬​🇹🇷​🇹🇭​🇳🇴​🇻🇳​🇦🇲​🇬🇪​🇩🇰​🇫🇮​🇬🇷​🇮🇱​🇲🇾​🇳🇱​🇸🇪​🇰🇪         |
-| **Higgs Audio 2**  | 3B                                        | ~9GB         | ✅ | ✅ | ❌ | ❌ | ❌ | Boson Higgs Audio 2 Community License | 3 multi-speaker, CUDA graphs (55+ tokens/sec)                                            | 🇺🇸​🇨🇳​🇩🇪​🇪🇸​🇰🇷                                                                           |
-| **IndexTTS-2**     | IndexTTS-2                                | ~4.7GB       | ✅ | ✅ | ❌ | ❌ | ❌ | bilibili Model Use License | Emotion Control: 8 vectors, Text as reference, Audio as reference                        | 🇺🇸​🇨🇳​🇯🇵                                                                                 |
-| **CosyVoice3**     | 0.5B, 0.5B-RL                             | ~5.4GB       | ✅ | ✅ | ✅ | ❌ | ❌ | Apache-2.0               | Paralinguistic tags                                                                      | 🇺🇸​🇨🇳​🇯🇵​🇰🇷                                                                              |
-| **Qwen3-TTS**      | 0.6B, 1.7B (CustomVoice/VoiceDesign/Base) | ~3-6GB       | ✅ | ✅ | ❌ | ✅ | ❌ | Apache-2.0               | Voice design, ASR (Automatic Speech Recognition)                                         | 🇺🇸​🇨🇳​🇩🇪​🇪🇸​🇫🇷​🇮🇹​🇯🇵​🇰🇷​🇷🇺​🇵🇹                                                            |
-| **Granite ASR**    | granite-4.0-1b-speech                     | ~4.6GB       | ❌ | ✅ | ❌ | ✅ | ❌ | Apache-2.0               | ASR (Automatic Speech Recognition), Custom timestamps/SRT via reused Qwen forced aligner, Speech translation (experimental) | 🇺🇸​🇩🇪​🇪🇸​🇫🇷​🇯🇵​🇵🇹                                                                        |
-| **Step Audio EditX** | 3B LLM + CosyVoice                        | ~7GB         | ✅ | ✅ | ❌ | ❌ | ❌ | Apache-2.0 (verify before commercial use) | Second Pass Speech Editing Node: 14 emotions, 32 speaking styles, Paralinguistic effects | 🇺🇸​🇨🇳​🇯🇵​🇰🇷                                                                              |
-| **Echo-TTS**       | echo-tts-base + fish-s1-dac-min           | ~5.3GB + ~1.8GB | ✅ | ✅ | ❌ | ❌ | ❌ | CC-BY-NC-SA-4.0          | Diffusion-based (~30s best), Force Speaker KV (speaker drift control)                    | 🇺🇸                                                                                       |
-| **MOSS-TTS**       | Local 1.7B, Delay 8B, TTSD 8B             | ~8.5GB tokenizer + ~6.1GB/17GB/18GB model | ✅ | ✅ | ❌ | ❌ | ✅ | Apache-2.0               | 20-language generation, Long-form generation (TTSD/Delay), Native 1-5 speaker TTSD dialogue, Duration token hint, Local/Delay/TTSD variants, Initial integrated LoRA training workflow (Delay 8B) | 🇺🇸​🇨🇳​🇩🇪​🇪🇸​🇫🇷​🇮🇹​🇯🇵​🇰🇷​🇷🇺​🇧🇷​🇵🇱​🇪🇬​🇹🇷​🇩🇰​🇬🇷​🇸🇪                                          |
-| **RVC**            | Community .pth                            | 100-300MB    | ❌ | ❌ | ✅ | ❌ | ✅ | MIT (framework); community models vary | Real-time VC, Integrated training workflow, Pitch shift (±14), 6 HuBERT models, Language-independent | 🇺🇸​🇨🇳​🇩🇪​🇪🇸​🇫🇷​🇮🇹​🇯🇵​🇰🇷​🇷🇺​🇧🇷​🇵🇱​🇮🇳​🇪🇬​🇹🇷​🇹🇭​🇳🇴​🇻🇳​🇦🇲​🇬🇪​🇩🇰​🇫🇮​🇬🇷​🇮🇱​🇲🇾​🇳🇱​🇸🇪​🇰🇪         |
+| Engine             | Isolation | Models                                    | Size         | TTS | SRT | VC  | ASR | Sound Effects | Training | License                  | Special Features                                                                         | Languages                                                                                |
+| ------------------ | --------- | ----------------------------------------- | ------------ | :-: | :-: | :-: | :-: | :-----------: | :------: | ------------------------ | ---------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------- |
+| **F5-TTS**         | Main      | Base, v1, E2TTS + 8 lang models           | ~1.2GB each  | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | CC-BY-NC-4.0             | Targeted Word/Speech Editing, Speed control                                              | 10                                                                                       |
+| **ChatterBox**     | Main      | EN, DE×3, IT, FR, RU, HY, KA, JA, KO, NO  | ~4.3GB       | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | MIT                      | Expressiveness slider                                                                    | 10                                                                                       |
+| **ChatterBox 23L** | Main      | v1, v2, v3, Vietnamese (Viterbox), Egyptian Arabic (oddadmix) | ~4.3GB       | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | MIT                      | V1, V2, and V3 official checkpoints, Emotion tokens (v2; currently ineffective), V3 skips the legacy alignment analyzer and trims the final token artifact | 25                                                                                       |
+| **VibeVoice**      | Shared    | 1.5B, 7B, KugelAudio-0 (7B), kugel-2 (7B), Hindi-1.5B/7B | 5.4GB / 18GB | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | MIT (research-only per model card) | 90-min long-form, Native 4-speaker (Base models), Multilingual (KugelAudio variants), 4-bit quantization | 27                                                                                       |
+| **Higgs Audio 2**  | Shared    | 3B                                        | ~9GB         | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | Boson Higgs Audio 2 Community License | 3 multi-speaker, CUDA graphs (55+ tokens/sec)                                            | 5                                                                                        |
+| **Higgs Audio v3** | Main      | 4B                                        | ~8GB         | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | Boson Higgs Audio v3 Research and Non-Commercial License | Native inline emotion/style/prosody/SFX tags, Zero-shot voice cloning, 100+ language support | 100+                                                                                     |
+| **IndexTTS 2 / 2.5** | Main      | IndexTTS-2, IndexTTS-2.5                  | ~4.7GB / ~5.49GB | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | bilibili Model Use License | Emotion Control: 8 vectors, Text as reference, Audio as reference, IndexTTS-2.5 official internal feature-duration scaling (not prosody planning), IndexTTS-2.5 pronunciation annotations | 5                                                                                        |
+| **CosyVoice3**     | Main      | 0.5B, 0.5B-RL                             | ~5.4GB       | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | Apache-2.0               | Paralinguistic tags                                                                      | 4                                                                                        |
+| **Qwen3-TTS**      | Shared    | 0.6B, 1.7B (CustomVoice/VoiceDesign/Base) | ~3-6GB       | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | Apache-2.0               | Voice design, ASR (Automatic Speech Recognition)                                         | 10                                                                                       |
+| **Granite ASR**    | Main      | granite-4.0-1b-speech, granite-speech-4.1-2b, granite-speech-4.1-2b-plus | ~4.6GB       | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | Apache-2.0               | Native speaker attribution / diarization (plus model variant), Native word-level timestamps (plus model variant), ASR (Automatic Speech Recognition), Custom timestamps/SRT via reused Qwen forced aligner, Speech translation (experimental), Optional forced aligner auto-routed through shared legacy T4 runtime | 6                                                                                        |
+| **Step Audio EditX** | Main      | 3B LLM + CosyVoice                        | ~7GB         | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | Apache-2.0 (verify before commercial use) | Second Pass Speech Editing Node: 14 emotions, 32 speaking styles, Paralinguistic effects, Selectable main, shared, or dedicated Python runtime (shared Transformers 4 runtime recommended) | 4                                                                                        |
+| **Echo-TTS**       | Main      | echo-tts-base + fish-s1-dac-min           | ~5.3GB + ~1.8GB | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | CC-BY-NC-SA-4.0          | Diffusion-based (~30s best), Force Speaker KV (speaker drift control)                    | 1                                                                                        |
+| **Fish Audio S2 Pro** | Main      | S2 Pro 4B / FP8                           | ~10.3GB / ~8.0GB | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | Fish Audio Research License | Free-form sub-word emotion/prosody tags, Native multi-speaker and multi-turn dialogue with dynamic speaker references, Zero-shot voice cloning, Optional per-segment custom character switching, Configurable 4K-32K native context with reduced KV-cache VRAM, Optional community FP8 weight-only checkpoint with BF16 activations, Optional on-the-fly BitsAndBytes INT8/NF4 for the official checkpoint | 80+ languages                                                                            |
+| **Dots TTS**       | Main      | dots.tts-base, dots.tts-soar, dots.tts-mf | ~6GB         | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | Apache-2.0               | Official auto language detect / language control, SOAR and MeanFlow distilled variants   | 19                                                                                       |
+| **DramaBox**       | Main      | DramaBox 3.3B                             | ~16.4GB      | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | LTX-2 Community License  | Expressive scene prompting and stage directions, Native and SRT-aware duration targeting, Official duration-aware long-form chunking with scene-prefix preservation, Optional 10-second zero-shot voice reference, CFG negative prompt with per-segment switching, Explicit generation/reference durations, CFG rescale control, and optional Perth watermark, Experimental staged and sequential strategies for lowering peak VRAM; official LTX FP8-cast storage, Optional official torch.compile path, Official audio-branch IC-LoRA training workflow | 1                                                                                        |
+| **OmniVoice**      | Main      | OmniVoice                                 | ~3.7GB       | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | Apache-2.0               | Inline non-verbal tags and pronunciation overrides, Reference-free voice design, 600+ language support, Upstream long-form chunk orchestration | 600+                                                                                     |
+| **MOSS-TTS**       | Main      | Local 1.7B, Delay 8B v1.5/1.0, LAION Voice Acting 8B community fine-tune, VoiceGenerator 1.7B, SoundEffect 8B v1, TTSD 8B | ~8.5GB tokenizer + ~6.1GB/17GB/18GB model | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | Apache-2.0               | Reference-free voice design with MOSS-VoiceGenerator, Native 1-5 speaker TTSD dialogue, 31-language generation with MOSS-TTS-v1.5, Optional LAION community 8B voice-acting fine-tune, Config-based discovery of compatible local MOSS full checkpoints, Prompt-only sound-effect generation with MOSS-SoundEffect v1, Long-form generation (TTSD/Delay), Duration token hint, Local/Delay/TTSD variants, Initial integrated LoRA training workflow (Delay 8B) | 24                                                                                       |
+| **MOSS-SoundEffect v2** | Main      | MOSS-SoundEffect-v2.0                     | ~11.2GB      | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | Apache-2.0               | Durations up to 30 seconds, Native negative prompting, CFG, flow shift, and diffusion-step controls, Prompt-only text-to-sound generation, 48 kHz mono output, Seeded generation | 2                                                                                        |
+| **RVC**            | Main      | Community .pth                            | 100-300MB    | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | MIT (framework); community models vary | Real-time VC, Integrated training workflow, Pitch shift (±14), 6 HuBERT models, Language-independent | Any                                                                                      |
+
+*Isolation column: `Main` runs in the main ComfyUI environment. `Shared` uses a shared secondary runtime reused by multiple engines. `Dedicated` uses an engine-specific secondary runtime.*

docs/FEATURE_COMPARISON.md

@@ -2,17 +2,22 @@
 
 ## Feature Comparison Matrix
 
-| Feature                      | F5-TTS | ChatterBox | ChatterBox 23L | VibeVoice | Higgs Audio 2 | IndexTTS-2 | CosyVoice3 | Qwen3-TTS | Granite ASR | Step Audio EditX | Echo-TTS | MOSS-TTS | RVC |
-|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
-| **TTS**                      | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ |
-| **SRT**                      | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-| **Voice Conversion**         | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
-| **ASR (Transcribe)**         | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
-| **Training**                 | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ |
-| **Voice Cloning**            | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ (Base model) | ❌ | ✅ | ✅ | ✅ | ⚠️ (needs training) |
-| **Native Multi-Speaker**     | ❌ | ❌ | ❌ | ✅ (Base only, Kugel uses fallback) | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (TTSD v1.0; 1-5 speakers) | ❌ |
-| **Emotion Control**          | ❌ | ❌ | ⚠️ (v2 tags - doesn't work) | ❌ | ⚠️ (via prompt) | ✅ (8 emotions) | ⚠️ (via instruct) | ⚠️ (via instruct) | ❌ | ✅ (14 emotions) | ❌ | ❌ | ❌ |
-| **Native Long-form**         | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (TTSD/Delay long-form; use chunk orchestration for very long inputs) | N/A |
-| **Community Finetunes**      | ✅ | ✅ | ✅ | ✅ KugelAudio, Hindi | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (LoRA adapter inference supported; initial integrated LoRA training support added for MOSS-TTS Delay 8B) | ✅ |
-| **VRAM Efficient**           | ✅ | ✅ | ✅ | ⚠️ (5-18GB) | ⚠️ (9GB) | ⚠️ (9-12GB) | ✅ (5.4GB) | ✅ (3-6GB) | ✅ (~4.6GB) | ⚠️ (7GB) | ⚠️ (~7GB total) | ⚠️ (Local 1.7B smaller; tokenizer is large) | ✅ |
-| **Speed/Performance**        | ✅ Very Fast | ✅ Fast | ✅ Fast | ⚠️ | ⚠️ | ⚠️ | ✅ Fast | ⚠️ | ⚠️ Moderate | ⚠️ | ✅ Fast (diffusion, realtime-capable) | ✅ Fast with CUDA/FlashAttention | ✅ Fast |
+| Feature                      | F5-TTS | ChatterBox | ChatterBox 23L | VibeVoice | Higgs Audio 2 | Higgs Audio v3 | IndexTTS 2 / 2.5 | CosyVoice3 | Qwen3-TTS | Granite ASR | Step Audio EditX | Echo-TTS | Fish Audio S2 Pro | Dots TTS | DramaBox | OmniVoice | MOSS-TTS | MOSS-SoundEffect v2 | RVC |
+|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
+| **TTS**                      | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+| **SRT**                      | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+| **Voice Conversion**         | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| **ASR (Transcribe)**         | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+| **Sound Effects**            | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ |
+| **Training**                 | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ✅ |
+| **Voice Cloning**            | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ (Base model) | ❌ | ✅ | ✅ | ✅ Reference audio plus exact transcript | ✅ | ✅ Optional reference audio; configurable 3-30 second window from the beginning of the source, default 10 seconds | ✅ | ✅ | ❌ | ⚠️ (needs training) |
+| **Reference Transcript†**    | **Required** | Not used | Not used | Not used | Optional | Optional | Not used | Conditional | Conditional | N/A | **Required** | Not used | **Required** | Optional | Not used | **Required** | Conditional | Not used | N/A |
+| **Native Multi-Speaker**     | ❌ | ❌ | ❌ | ✅ (Base only, Kugel uses fallback) | ✅ | ❌ | ❌ | ❌ | ❌ | ⚠️ (Plus variant speaker attribution / diarization) | ❌ | ❌ | ✅ Native dialogue is default; optional custom mode generates each [Character] segment independently as local speaker 0 | ❌ | ❌ | ❌ | ✅ (TTSD v1.0; 1-5 speakers) | ❌ | ❌ |
+| **Emotion Control**          | ❌ | ❌ | ⚠️ (v2 tags - doesn't work) | ❌ | ⚠️ (via prompt) | ✅ (native inline tags) | ✅ (8 emotions) | ⚠️ (via instruct) | ⚠️ (via instruct) | ❌ | ✅ (14 emotions) | ❌ | ✅ Free-form inline natural-language tags | ❌ | ✅ Natural-language scene prompt and stage directions | ⚠️ (voice-design instruct + inline non-verbal tags) | ✅ (MOSS-VoiceGenerator instruction-conditioned voice design) | ❌ | ❌ |
+| **Native Long-form**         | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ Configurable 4K-32K native context; suite text chunking is bypassed | ❌ | ✅ Official duration-aware quote-group chunking; ~37s target / 45s cap | ✅ (uses upstream audio_chunk_duration / audio_chunk_threshold orchestration; bypasses suite char-based chunk splitting) | ✅ (TTSD/Delay long-form; use chunk orchestration for very long inputs) | ❌ | N/A |
+| **Community Finetunes**      | ✅ | ✅ | ✅ | ✅ KugelAudio, Hindi | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ Official audio-branch IC-LoRA adapters can be trained and loaded | ❌ | ✅ (Compatible full local checkpoints, LAION Voice Acting 8B auto-download, and LoRA adapter inference/training supported) | ❌ | ✅ |
+| **VRAM Efficient**           | ✅ | ✅ | ✅ | ⚠️ (5-18GB) | ⚠️ (9GB) | ⚠️ (~8-10GB) | ⚠️ (9-12GB) | ✅ (5.4GB) | ✅ (3-6GB) | ✅ (~4.6GB) | ⚠️ (7GB) | ⚠️ (~7GB total) | ⚠️ 8K context measured at ~15.2GB BF16, ~11.2GB FP8, ~11.2GB BNB INT8, or ~8.9GB BNB NF4; BF16 codec and activations; BNB is a load-time option for the official checkpoint | ⚠️ (main env works; 2B-class model is not lightweight) | ✅ Fast mode is ~24GB; experimental FP8, staged, and sequential options can reduce VRAM, but practical peak usage varies by environment and loaded components | ⚠️ (main env works; 2B-class model is not lightweight) | ⚠️ (Local 1.7B smaller; tokenizer is large) | ⚠️ (runs in the main ComfyUI environment but remains GPU-heavy) | ✅ |
+| **Speed/Performance**        | ✅ Very Fast | ✅ Fast | ✅ Fast | ⚠️ | ⚠️ | ⚠️ CUDA recommended | ⚠️ | ✅ Fast | ⚠️ | ⚠️ Moderate | ⚠️ | ✅ Fast (diffusion, realtime-capable) | ✅ Main-environment subprocess with reliable teardown; local compile measurements: ~40 it/s BF16, ~11.8 it/s NF4 at ~8.9GB VRAM, and ~3.7 it/s INT8 at ~11.2GB VRAM; quality comparison pending | ⚠️ Moderate; mf variant is faster | ⚠️ Fast favors repeated generation; staged reloads temporary components; sequential also transfers the transformer; optional block-level torch.compile accelerates repeat denoising | ✅ Fast; upstream reports sub-realtime RTF | ✅ Fast with CUDA/FlashAttention | ⚠️ (100 diffusion steps by default) | ✅ Fast |
+| **No Narrator Required**     | ❌ | ✅ (default speaker) | ✅ (default speaker) | ✅ (zero-shot / default speaker) | ✅ (basic TTS if no narrator/reference is provided) | ✅ (zero-shot) | ❌ | ✅ (cross-lingual or instruct mode) | ✅ (Base default voice or CustomVoice presets) | N/A | ❌ | ❌ | ✅ Reference audio is optional | ✅ (default speaker) | ✅ Voice reference is optional | ✅ (native default voice; instruct also works without narrator) | ✅ (direct TTS and prompt-only generation) | ❌ | N/A |
+
+† **Reference Transcript:** Conditional means the transcript is required only for the specific mode: CosyVoice3 zero-shot, Qwen3-TTS full Base cloning, or MOSS-TTSD cloned-speaker dialogue. Higgs Audio 2, Higgs Audio v3, and Dots TTS accept matching text when provided but do not require it.

docs/FISH_AUDIO_S2_TAGS.md

@@ -0,0 +1,35 @@
+# Fish Audio S2 Pro Inline Tags
+
+Use the suite's public angle-bracket syntax for Fish's free-form instructions:
+
+```text
+<whisper in small voice>Hello. <professional broadcast tone>Good evening.
+```
+
+The Fish adapter converts these to native `[...]` instructions only when text reaches the Fish engine. Do not write Fish tags with square brackets in suite text because `[Character]` is reserved for character switching.
+
+Fish also supports normal suite tags such as `[pause:500ms]`, character switching, and per-segment parameters. Those are processed by the suite before Fish inference.
+
+## Language Prompting
+
+Fish has no native language dropdown or parameter. The engine can instead prepend a natural inline instruction from the suite's resolved segment language:
+
+- `language_prompting = Auto Inline Tag`: non-English resolved languages become tags such as `<German>` or `<French>`
+- `language_prompting = Off`: no automatic language instruction is added
+
+English is only added when the user explicitly requested it with a language tag such as `[en:Bob]` or `[English:Bob]`. Implicit/default English stays untagged.
+
+## Character Switching Modes
+
+The engine defaults to `Native Multi-Speaker`. All `[Character]` turns in a generated block are sent through one Fish dialogue request, preserving native multi-turn context and long-form behavior.
+
+Select `Custom Character Switching` to generate every parsed character segment independently. Each call uses only that character's reference and is remapped to Fish speaker 0. This can reduce speaker leakage, but it requires more calls and does not preserve context between character turns. SRT subtitle boundaries and timing are preserved in both modes.
+
+The engine UI separates checkpoint choice from load-time quantization:
+
+- `model_variant`: `s2-pro` or the separate `s2-pro-fp8` checkpoint
+- `quantization`: `none`, `bnb_int8`, or `bnb_nf4` for on-the-fly quantization of the official `s2-pro` checkpoint
+
+The BNB options reuse the official files and do not download a second model copy.
+
+The S2 Pro weights use the Fish Audio Research License: research and non-commercial use are allowed; commercial use requires a separate Fish Audio license.

docs/HIGGS_AUDIO_V3_INLINE_TAGS.md

@@ -0,0 +1,78 @@
+# Higgs Audio v3 Inline Tags
+
+Higgs Audio v3 uses native inline control tokens in the text prompt. Type these tags directly in **TTS Text** or **TTS SRT** when using the Higgs Audio v3 engine.
+
+## Syntax
+
+Canonical native tags use:
+
+```text
+<|category:value|>
+```
+
+TTS Audio Suite also accepts a convenience alias form without the bars:
+
+```text
+<category:value>
+```
+
+Aliases are normalized internally to the canonical Higgs syntax. The editor inserts the canonical `<|...|>` form on purpose.
+
+Examples:
+
+```text
+<|emotion:amusement|>Wait, that was actually funny.
+<|style:whispering|>Keep your voice down.
+I need a second.<|prosody:pause|> Okay, continue.
+That was perfect. <|sfx:laughter|>Haha, absolutely perfect.
+```
+
+## Supported Categories
+
+Emotion:
+
+```text
+elation, amusement, enthusiasm, determination, pride, contentment, affection,
+relief, contemplation, confusion, surprise, awe, longing, arousal, anger, fear,
+disgust, bitterness, sadness, shame, helplessness
+```
+
+Style:
+
+```text
+singing, shouting, whispering
+```
+
+Prosody:
+
+```text
+speed_very_slow, speed_slow, speed_fast, speed_very_fast,
+pitch_low, pitch_high, pause, long_pause,
+expressive_high, expressive_low
+```
+
+Sound effects:
+
+```text
+cough, laughter, crying, screaming, burping, humming, sigh, sniff, sneeze
+```
+
+## Placement Rules
+
+- Put delivery-wide tags like emotion, style, speed, pitch, and expressiveness at the start of the sentence or turn.
+- Put positional tags like `<|prosody:pause|>`, `<|prosody:long_pause|>`, and `<|sfx:...|>` exactly where they should happen.
+- Pair SFX tags with written sound text, for example `<|sfx:laughter|>Haha` or `<|sfx:sigh|>Ahh`.
+- Canonical Higgs syntax is `<|...|>`. Alias input like `<emotion:amusement>` is accepted by TTS Audio Suite for convenience, but the official form stays `<|emotion:amusement|>`.
+- Higgs Audio v3 does not use Step Audio EditX second-pass inline processing here. Use Higgs native tags instead of Step semantics when targeting Higgs.
+
+## Character Voices
+
+Higgs Audio v3 supports reference audio cloning. A transcript of the reference audio strongly improves cloning quality. Character voice `.txt` files are passed as reference transcripts when available.
+
+## Multiline Editor
+
+The multiline TTS tag editor has an `Inline Tags` panel with a dedicated `Higgs Audio v3` mode. That panel inserts canonical native Higgs tags directly.
+
+## License
+
+Higgs Audio v3 is released under the Boson Higgs Audio v3 Research and Non-Commercial License. Production, hosted API, or revenue-generating use requires a separate commercial license from Boson AI.

docs/INLINE_EDIT_TAGS_USER_GUIDE.md

@@ -1,4 +1,15 @@
-# Inline Edit Tags - User Guide
+# Step Audio EditX Inline Tags - User Guide
+
+## Scope
+
+This guide is specifically for **Step Audio EditX** inline post-processing tags.
+
+The multiline TTS tag editor now has an engine-aware `Inline Tags` panel with separate modes for:
+- `Step Audio EditX`
+- `Higgs Audio v3`
+- `CosyVoice3`
+
+Use the Step mode when you want post-process editing tags. Use the Higgs or Cosy modes when you want those engines' native generation tags.
 
 ## 💡 Inline vs Manual Node Workflow
 
@@ -15,7 +26,7 @@
 
 ## Overview
 
-Inline edit tags allow you to apply **Step Audio EditX** post-processing effects to TTS-generated audio from ANY engine (F5-TTS, ChatterBox, Higgs Audio, VibeVoice, etc.). These tags are automatically stripped before TTS generation and applied as a second processing pass.
+Inline edit tags allow you to apply **Step Audio EditX** post-processing effects to TTS-generated audio from engines that support the Step inline workflow. These tags are automatically stripped before TTS generation and applied as a second processing pass.
 
 **Example:**
 ```
@@ -661,6 +672,10 @@ A: Just remove them from your text. No special setting needed.
 
 ## Support
 
+Related guides:
+- `docs/HIGGS_AUDIO_V3_INLINE_TAGS.md` for Higgs Audio v3 native tags
+- `docs/COSYVOICE3_TAGS_GUIDE.md` for CosyVoice3 native tags
+
 For issues, feature requests, or questions:
 - GitHub Issues: https://github.com/diodiogod/TTS-Audio-Suite/issues
 - Check console output for detailed processing information

docs/IndexTTS2_Emotion_Control_Guide.md

@@ -12,15 +12,27 @@ IndexTTS-2 supports multiple emotion control methods that can be combined for so
 - **Text Emotion**: AI-powered QwenEmotion analysis from text descriptions with dynamic templates
 - **Character Tag Emotions**: Per-character emotion control using `[Character:emotion_ref]` syntax
 
-## Emotion Control Priority
-
-You can only connect to the Engine node one source of control emotion: Either audio, text, or vectors.
-
-When using tags on the text iself, **Character tag emotions** (highest priority) - `[Alice:angry_bob]` overrides all other emotion control settings for that character segment
+## Emotion Control Inputs and Blending
+
+The IndexTTS-2 Engine has two emotion inputs:
+
+- **`emotion_control`**: vector controls or Qwen text emotion
+- **`emotion_audio`**: an audio reference from an AUDIO or Character Voices node
+
+Both inputs may be connected at the same time. Text emotion is analyzed into an
+8-value vector; audio emotion remains an audio-derived conditioning signal. The
+engine blends the two signals in its latent emotion-conditioning space rather
+than converting the audio into the eight visible vector values.
+
+For backward compatibility, the original `emotion_control` socket still accepts
+legacy audio connections, but new workflows should use `emotion_audio` for
+audio references. A character tag such as `[Alice:angry_bob]` supplies a
+segment-local audio reference and can also be combined with vector/text emotion
+for that segment.
 
 ## Method 1: Direct Audio Reference
 
-Connect any audio file directly to the IndexTTS-2 Engine's `emotion_control` input.
+Connect any audio file directly to the IndexTTS-2 Engine's `emotion_audio` input.
 
 **How it works:**
 
@@ -37,7 +49,7 @@ Connect any audio file directly to the IndexTTS-2 Engine's `emotion_control` inp
 **Example:**
 
 ```
-AUDIO node → IndexTTS-2 Engine (emotion_control)
+AUDIO node → IndexTTS-2 Engine (emotion_audio)
 ```
 
 ## Method 2: Character Voices Audio Reference
@@ -48,7 +60,7 @@ Use the `opt_narrator` output from the 🎭 Character Voices node as an emotion
 
 1. Add a 🎭 Character Voices node
 2. Select a voice with the desired emotional expression
-3. Connect `opt_narrator` output to IndexTTS-2 Engine `emotion_control` input
+3. Connect `opt_narrator` output to IndexTTS-2 Engine `emotion_audio` input
 
 **Advantages:**
 
@@ -59,12 +71,13 @@ Use the `opt_narrator` output from the 🎭 Character Voices node as an emotion
 **Example workflow:**
 
 ```
-🎭 Character Voices (David_Attenborough) → opt_narrator → IndexTTS-2 Engine (emotion_control)
+🎭 Character Voices (David_Attenborough) → opt_narrator → IndexTTS-2 Engine (emotion_audio)
 ```
 
-## Method 3: Emotion Vectors
-
-Use the 🌈 IndexTTS-2 Emotion Vectors node for precise manual control over 8 different emotions.
+## Method 3: Emotion Vectors
+
+Use the 🌈 IndexTTS-2 Emotion Vectors node for precise manual control over 8 different emotions.
+Connect its `emotion_control` output to the IndexTTS-2 Engine's `emotion_control` input.
 
 **Available emotions:**
 
@@ -84,9 +97,10 @@ Use the 🌈 IndexTTS-2 Emotion Vectors node for precise manual control over 8 d
 - Start with single emotions, then experiment with combinations
 - Use the `random` buttom to get a completely random emotion pattern. Might be too strong.
 
-## Method 4: Text Emotion (Dynamic Analysis)
-
-Use the 🌈 IndexTTS-2 Text Emotion node for AI-powered emotion analysis with dynamic templates.
+## Method 4: Text Emotion (Dynamic Analysis)
+
+Use the 🌈 IndexTTS-2 Text Emotion node for AI-powered emotion analysis with dynamic templates.
+Connect its `emotion_control` output to the IndexTTS-2 Engine's `emotion_control` input.
 
 ### Static Text Emotion
 
@@ -124,11 +138,104 @@ Analysis: "Worried parent speaking: Where have you been?"
 Result: Anxious, concerned vocal expression
 ```
 
----
-
-## Character Tag Emotion Control
-
-Control emotions per character using inline tags in your text: `[Character:emotion_ref]`
+---
+
+## Combining Audio with Vectors or Text
+
+Connect both emotion sources when you want an audio performance to provide the
+base delivery while vectors or Qwen text analysis add a targeted emotional
+direction:
+
+```text
+🎭 Character Voices (opt_narrator) ──→ emotion_audio
+🌈 Emotion Vectors or Text Emotion ──→ emotion_control
+                                      IndexTTS-2 Engine
+```
+
+`emotion_alpha` is the shared overall emotion-intensity control. The audio
+reference and vector/text signal are blended during IndexTTS-2 conditioning;
+they are not generated as two separate voices and mixed afterward.
+
+For example, an audio reference can provide a natural speaking style while a
+`[sad:+0.2|calm:-0.1]` inline adjustment adds a restrained sadness to one
+segment. A Qwen text preset can be used the same way.
+
+Character audio and inline emotion parameters can share one tag:
+
+```text
+[Bob:br_ivan_raiva3|sad:+0.25|calm:-0.10] Bob speaks with a restrained overlay.
+[Bob:br_ivan_raiva3|emotion:"quiet grief masking frustration"] Bob uses Qwen text emotion too.
+```
+
+---
+
+## Inline Emotion Switching
+
+Numeric emotion tags can replace or adjust the vector for one text segment:
+
+```text
+[sad:0.7|calm:0.2] This uses explicit absolute values.
+[sad:+0.3|calm:-0.2] This modifies the connected vector.
+[vector:0,0,0.7,0,0,0.4,0,0.2] This supplies all eight absolute values.
+[vector:+0,+0,+0.3,+0,+0,+0,+0,-0.2] This supplies eight deltas.
+```
+
+The full-vector order is `happy, angry, sad, afraid, disgusted, melancholic,
+surprised, calm`. Unsigned named values are absolute; `+` and `-` named values
+are relative. A full vector is relative only when every value carries a sign.
+
+Text-emotion controls support saved presets and quoted descriptions:
+
+```text
+[emotion:restrained_anger] Text using a saved preset.
+[emotion:"Restrained anger masking disappointment"] Direct text control.
+[emotion:"Analyze this delivery as nervous anticipation: {seg}"] Dynamic control.
+```
+
+Inline controls override connected global vector/text values for that segment.
+A character audio emotion reference such as `[Alice:sad_reference]` replaces the
+global audio reference for that segment, but it can still blend with the
+segment's vector/text control. Inline settings revert at the next segment and do
+not mutate the connected vector.
+
+The TTS Tag Editor provides the same interactive radar used by the IndexTTS-2
+Emotion Vectors node. Click an existing numeric emotion tag to open its radar
+as a contextual popover beside the tag. Create tags from **Inline Tags →
+IndexTTS-2**, and use **Manage Emotion Presets** in its Text Emotion section for
+the preset library. Text and vector presets are stored in
+`models/TTS/IndexTTS/emotion_presets.json`.
+
+Radar changes appear in the editor text immediately. Intermediate drag/input
+states are not added to undo history: closing the popover commits one undoable
+change, while Cancel or Escape restores the tag exactly as it was when opened.
+
+The editor's **Inline Tags** tab also includes an **IndexTTS-2** engine panel for
+inserting full absolute/delta vectors, named emotion values, saved text presets,
+quoted descriptions, and dynamic descriptions containing `{seg}`.
+
+Emotion controls can be composed directly on character tags: place the caret on
+`[Bob:audio_reference]` and add a vector or text emotion to append it as a pipe
+parameter.
+
+The named-emotion panel includes a magnitude slider and a press-drag-release
+radial picker: direction chooses the emotion and distance chooses its value.
+The operation dropdown determines whether that value is absolute, a positive
+delta, or a negative delta. Saved text and vector presets refresh in the sidebar
+immediately after they are changed in the preset manager.
+
+Clicking a saved `[emotion:preset_name]` tag in the editor opens a small anchored
+preset dropdown, allowing that line's preset to be swapped without opening the
+full manager. Adding an emotion control while the caret is inside a pure emotion
+tag replaces that tag; when the caret is inside a character/audio tag, the
+editor appends or updates the emotion parameter after the existing `|` fields.
+
+Named tags remain readable while only some emotions are active. If radar editing
+activates all eight emotions, the editor automatically converts the result to the
+shorter ordered `[vector:...]` form.
+
+## Character Tag Emotion Control
+
+Control emotions per character using inline tags in your text: `[Character:emotion_ref]`
 
 **Syntax:**
 
@@ -145,8 +252,9 @@ Control emotions per character using inline tags in your text: `[Character:emoti
 
 ```
 Hello everyone! [Alice:happy_sarah] I'm so excited to be here today!
-[Bob:angry_tom] That's completely unacceptable behavior.
-[Narrator:David] Meanwhile, in a distant galaxy...
+[Bob:angry_tom] That's completely unacceptable behavior.
+[Narrator:David] Meanwhile, in a distant galaxy...
+[Bob:br_ivan_raiva3|sad:+0.25] Bob uses an audio reference plus a vector delta.
 
 *assuming happy_sarah, angry_tom and David are alias or character voices in yout folder with that name
 ```
@@ -155,11 +263,13 @@ Hello everyone! [Alice:happy_sarah] I'm so excited to be here today!
 
 
 
-**Character tag priority:**
-
-- Character tags override ALL other emotion settings for that specific character
-- Other characters use global emotion settings
-- Allows mixing different emotions in the same audio
+**Character tag behavior:**
+
+- Character tags select the speaker and can provide a segment-local audio emotion reference
+- A segment-local audio reference replaces the global audio reference for that segment
+- Global or inline vector/text controls can still blend with that audio reference
+- Other characters use the global audio and vector/text settings
+- Allows mixing different emotion sources in the same audio
 
 ## Emotion Alpha Control
 
@@ -211,4 +321,4 @@ Welcome to our show! [Bob:serious_narrator] But first, a serious announcement.
 
 ---
 
-This comprehensive emotion control system gives you unprecedented flexibility in creating expressive, emotionally rich TTS audio for any application.
+This comprehensive emotion control system gives you unprecedented flexibility in creating expressive, emotionally rich TTS audio for any application.

docs/LANGUAGE_SUPPORT.md

@@ -2,41 +2,118 @@
 
 ## Language Support by Engine
 
-| Language       | Code | F5-TTS | ChatterBox | ChatterBox 23L | VibeVoice | Higgs Audio 2 | IndexTTS-2 | CosyVoice3 | Qwen3-TTS | Granite ASR | Step Audio EditX | Echo-TTS | MOSS-TTS | RVC |
-|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
-| 🇺🇸 **English** | EN   | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-| 🇨🇳 **Chinese** | ZH   | ❌ | ❌ | ✅ | ✅ | ✅ (Mandarin) | ✅ | ✅ + 18 dialects | ✅ | ❌ | ✅ (Mandarin + Sichuanese, Cantonese) | ❌ | ✅ | ✅ |
-| 🇩🇪 **German**  | DE   | ✅ | ✅ (×3) | ✅ | ✅ (Kugel) | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ |
-| 🇪🇸 **Spanish** | ES   | ✅ | ❌ | ✅ | ✅ (Kugel) | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ |
-| 🇫🇷 **French**  | FR   | ✅ | ✅ | ✅ | ✅ (Kugel) | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ |
-| 🇮🇹 **Italian** | IT   | ✅ | ✅ | ✅ | ✅ (Kugel) | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ |
-| 🇯🇵 **Japanese** | JA   | ✅ | ✅ | ✅ | ✅ (Kugel) | ❌ | ✅ ? | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
-| 🇰🇷 **Korean**  | KO   | ❌ | ✅ | ✅ | ✅ (Kugel) | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ |
-| 🇷🇺 **Russian** | RU   | ❌ | ✅ | ✅ | ✅ (Kugel) | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ |
-| 🇧🇷 **Portuguese** | PT   | ✅ (BR) | ❌ | ✅ | ✅ (Kugel) | ❌ | ❌ | ❌ | ✅ (EU/BR*) | ✅ | ❌ | ❌ | ✅ | ✅ |
-| 🇵🇱 **Polish**  | PL   | ✅ | ❌ | ✅ | ✅ (Kugel) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ |
-| 🇮🇳 **Hindi**   | HI   | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
-| 🇪🇬 **Arabic**  | AR   | ❌ | ❌ | ✅ (Egyptian) | ✅ (Kugel) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ |
-| 🇹🇷 **Turkish** | TR   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ |
-| 🇹🇭 **Thai**    | TH   | ✅ | ❌ | ✅ | ✅ (Kugel) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
-| 🇳🇴 **Norwegian** | NO   | ❌ | ✅ | ✅ | ✅ (Kugel) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
-| 🇻🇳 **Vietnamese** | VI   | ❌ | ❌ | ✅ (Viterbox) | ✅ (Kugel) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
-| 🇦🇲 **Armenian** | HY   | ❌ | ✅ | ❌ | ✅ (Kugel) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
-| 🇬🇪 **Georgian** | KA   | ❌ | ✅ | ❌ | ✅ (Kugel) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
-| 🇩🇰 **Danish**  | DA   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ |
-| 🇫🇮 **Finnish** | FI   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
-| 🇬🇷 **Greek**   | EL   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ |
-| 🇮🇱 **Hebrew**  | HE   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
-| 🇲🇾 **Malay**   | MS   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
-| 🇳🇱 **Dutch**   | NL   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
-| 🇸🇪 **Swedish** | SV   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ |
-| 🇰🇪 **Swahili** | SW   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| Language       | Code | F5-TTS | ChatterBox | ChatterBox 23L | VibeVoice | Higgs Audio 2 | Higgs Audio v3 | IndexTTS 2 / 2.5 | CosyVoice3 | Qwen3-TTS | Granite ASR | Step Audio EditX | Echo-TTS | Fish Audio S2 Pro | Dots TTS | DramaBox | OmniVoice | MOSS-TTS | MOSS-SoundEffect v2 | RVC |
+|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
+| 🇺🇸 **English** | EN   | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ Tier 1 | ✅ | ✅ Official model is English-only | ✅ | ✅ | ✅ Officially demonstrated prompt language | ✅ |
+| 🇨🇳 **Chinese** | ZH   | ❌ | ❌ | ✅ | ✅ | ✅ (Mandarin) | ✅ | ✅ | ✅ + 18 dialects | ✅ | ❌ | ✅ (Mandarin + Sichuanese, Cantonese) | ❌ | ✅ Tier 1 | ✅ (Mandarin; official also exposes YUE separately outside this matrix) | ❌ | ✅ | ✅ | ✅ Officially demonstrated prompt language | ✅ |
+| 🇩🇪 **German**  | DE   | ✅ | ✅ (×3) | ✅ | ✅ (Kugel) | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ Tier 2 | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ |
+| 🇪🇸 **Spanish** | ES   | ✅ | ❌ | ✅ | ✅ (Kugel) | ✅ | ✅ | ✅ IndexTTS-2.5 | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ Tier 2 | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ |
+| 🇫🇷 **French**  | FR   | ✅ | ✅ | ✅ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ Tier 2 | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ |
+| 🇮🇹 **Italian** | IT   | ✅ | ✅ | ✅ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ |
+| 🇯🇵 **Japanese** | JA   | ✅ | ✅ | ✅ | ✅ (Kugel) | ❌ | ✅ | ✅ IndexTTS-2.5 | ✅ | ✅ | ✅ (4.1-2b supports Japanese; plus variant does not) | ✅ | ❌ | ✅ Tier 1 | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ |
+| 🇰🇷 **Korean**  | KO   | ❌ | ✅ | ✅ | ✅ (Kugel) | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ Tier 2 | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ |
+| 🇷🇺 **Russian** | RU   | ❌ | ✅ | ✅ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ Tier 2 | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ |
+| 🇧🇷 **Portuguese** | PT   | ✅ (BR) | ❌ | ✅ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ✅ (EU/BR*) | ✅ | ❌ | ❌ | ✅ Tier 2 | ✅ (Official PT tag is generic; upstream does not expose separate PT-BR/PT-PT tags and it may lean more European Portuguese than Brazilian Portuguese) | ❌ | ✅ (generic PT; official language space is much broader than this matrix) | ✅ | ❌ | ✅ |
+| 🇵🇱 **Polish**  | PL   | ✅ | ❌ | ✅ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ |
+| 🇮🇳 **Hindi**   | HI   | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ (v1.5) | ❌ | ✅ |
+| 🇪🇬 **Arabic**  | AR   | ❌ | ❌ | ✅ (Egyptian) | ✅ (Kugel) | ❌ | ✅ | ✅ IndexTTS-2.5 | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ Tier 2 | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ |
+| 🇹🇷 **Turkish** | TR   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ |
+| 🇹🇭 **Thai**    | TH   | ✅ | ❌ | ✅ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ (v1.5) | ❌ | ✅ |
+| 🇳🇴 **Norwegian** | NO   | ❌ | ✅ | ✅ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ |
+| 🇻🇳 **Vietnamese** | VI   | ❌ | ❌ | ✅ (Viterbox) | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ (v1.5) | ❌ | ✅ |
+| 🇦🇲 **Armenian** | HY   | ❌ | ✅ | ❌ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ |
+| 🇬🇪 **Georgian** | KA   | ❌ | ✅ | ❌ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ |
+| 🇩🇰 **Danish**  | DA   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ |
+| 🇫🇮 **Finnish** | FI   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ (v1.5) | ❌ | ✅ |
+| 🇬🇷 **Greek**   | EL   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ |
+| 🇮🇱 **Hebrew**  | HE   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ (v1.5) | ❌ | ✅ |
+| 🇲🇾 **Malay**   | MS   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ (v1.5) | ❌ | ✅ |
+| 🇳🇱 **Dutch**   | NL   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ (v1.5) | ❌ | ✅ |
+| 🇸🇪 **Swedish** | SV   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ |
+| 🇰🇪 **Swahili** | SW   | ❌ | ❌ | ✅ | ✅ (Kugel) | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ (v1.5) | ❌ | ✅ |
+| 🇿🇦 **Afrikaans** | AF   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇦🇱 **Albanian** | SQ   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (usable) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇮🇳 **Assamese** | AS   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇪🇸 **Asturian** | AST  | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇦🇿 **Azerbaijani** | AZ   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇷🇺 **Bashkir** | BA   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇪🇸 **Basque**  | EU   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇧🇾 **Belarusian** | BE   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇧🇩 **Bengali** | BN   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇧🇦 **Bosnian** | BS   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇧🇬 **Bulgarian** | BG   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇪🇸 **Catalan** | CA   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇵🇭 **Cebuano** | CEB  | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇮🇶 **Central Kurdish** | CKB  | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇭🇷 **Croatian** | HR   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇨🇿 **Czech**   | CS   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇷🇺 **Eastern Mari** | MHR  | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🌍 **Esperanto** | EO   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇪🇪 **Estonian** | ET   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇪🇸 **Galician** | GL   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇮🇳 **Gujarati** | GU   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇭🇹 **Haitian Creole** | HT   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇳🇬 **Hausa**   | HA   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇭🇺 **Hungarian** | HU   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇮🇩 **Indonesian** | ID   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇮🇩 **Javanese** | JV   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇮🇳 **Kannada** | KN   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇰🇿 **Kazakh**  | KK   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇷🇼 **Kinyarwanda** | RW   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇰🇬 **Kyrgyz**  | KY   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇱🇻 **Latvian** | LV   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇨🇩 **Lingala** | LN   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇱🇹 **Lithuanian** | LT   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇰🇪 **Luo**     | LUO  | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇲🇰 **Macedonian** | MK   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇮🇳 **Malayalam** | ML   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇲🇹 **Maltese** | MT   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇳🇿 **Māori**   | MI   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇮🇳 **Marathi** | MR   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇲🇳 **Mongolian** | MN   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇳🇵 **Nepali**  | NE   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇫🇷 **Occitan** | OC   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇮🇷 **Persian** | FA   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇷🇴 **Romanian** | RO   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇿🇦 **Sepedi**  | NSO  | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇷🇸 **Serbian** | SR   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇿🇼 **Shona**   | SN   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇸🇰 **Slovak**  | SK   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇸🇮 **Slovene** | SL   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇵🇭 **Tagalog** | TL   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇹🇯 **Tajik**   | TG   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇮🇳 **Tamil**   | TA   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇮🇳 **Telugu**  | TE   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇺🇦 **Ukrainian** | UK   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇵🇰 **Urdu**    | UR   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇨🇳 **Uyghur**  | UG   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇺🇿 **Uzbek**   | UZ   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇿🇦 **Xhosa**   | XH   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇿🇦 **Zulu**    | ZU   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (polished) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇲🇼 **Chichewa/Nyanja** | NY   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (usable) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇮🇳 **Eastern Punjabi** | PA   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (usable) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇺🇬 **Ganda**   | LG   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (usable) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇮🇸 **Icelandic** | IS   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (usable) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇮🇪 **Irish**   | GA   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (usable) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇩🇿 **Kabyle**  | KAB  | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (usable) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇨🇻 **Kabuverdianu** | KEA  | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (usable) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇰🇪 **Kamba**   | KAM  | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (usable) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇻🇦 **Latin**   | LA   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (usable) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇱🇺 **Luxembourgish** | LB   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (usable) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇪🇹 **Oromo**   | OM   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (usable) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇦🇫 **Pashto**  | PS   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (usable) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇵🇰 **Sindhi**  | SD   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (usable) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇸🇴 **Somali**  | SO   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (usable) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🇦🇴 **Umbundu** | UMB  | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (usable) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+| 🏴 **Welsh**    | CY   | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ (usable) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
 
 **Notes:**
 
 - **CosyVoice3 Chinese**: Includes 18+ dialects (Cantonese, Sichuan, Dongbei, Shanghai, etc.)
 - **Higgs Audio 2**: Trained on EN, ZH (Mandarin), KO, DE, ES (English majority) - 10M hours AudioVerse dataset
+- **Higgs Audio v3**: Official model card lists 102 languages total: 85 polished-production languages and 17 usable but less polished languages.
 - **IndexTTS-2**: Trained on 55K+ hours - ZH, EN, JA primary.
 - **MOSS-TTS**: Official list includes ZH, EN, DE, ES, FR, JA, IT, HU, KO, RU, FA, AR, PL, PT, CS, DA, SV, EL, TR; HU/FA/CS are not separate columns in this matrix.
-- **RVC**: Language-independent (work as post-processors VC with any TTS output)
+- **OmniVoice**: Official model supports 600+ languages; this matrix only shows the suite's comparison subset.
+- **RVC**: Language-agnostic voice conversion/post-processing; it is marked supported for every language row.
 - Qwen3-TTS: Supports pt-BR only with instruction (voice design or harcoded voices). Base can't do instructions, so it will always ouput pt-PT.

docs/MODEL_DOWNLOAD_SOURCES.md

@@ -38,7 +38,7 @@ Use this as the canonical list of model repositories/links for offline setup.
 
 | Component | Source | Size | Auto-Download | Notes |
 |---|---|---|---|---|
-| Official 23-Lang (v1/v2) | [ResembleAI/chatterbox](https://huggingface.co/ResembleAI/chatterbox) | ~4.3GB | ✅ | v1 + v2 files and tokenizer |
+| Official 23-Lang (v1/v2/v3) | [ResembleAI/chatterbox](https://huggingface.co/ResembleAI/chatterbox) | ~4.3GB | ✅ | v1 + v2 + v3 T3 checkpoints with shared tokenizer, voice encoder, and S3Gen |
 | Russian stress dictionary (Russian only) | [Vuizur/add-stress-to-epub release](https://github.com/Vuizur/add-stress-to-epub/releases/download/v1.0.1/russian_dict.zip) | ~1.5GB | ✅ | Auxiliary Official 23-Lang Russian stress-labeling data; downloads on demand only when Russian stress support is used |
 | Vietnamese (Viterbox) | [dolly-vn/viterbox](https://huggingface.co/dolly-vn/viterbox) | ~4.3GB | ✅ | Vietnamese community finetune used by downloader |
 | Egyptian Arabic (oddadmix) | [oddadmix/chatterbox-egyptian-v0](https://huggingface.co/oddadmix/chatterbox-egyptian-v0) | ~4.3GB | ✅ | Egyptian Arabic community finetune (architecture v2) |
@@ -59,11 +59,18 @@ Use this as the canonical list of model repositories/links for offline setup.
 | higgs-audio-v2-3B | [bosonai/higgs-audio-v2-generation-3B-base](https://huggingface.co/bosonai/higgs-audio-v2-generation-3B-base) | ~9GB | ✅ | Generation model |
 | Audio tokenizer | [bosonai/higgs-audio-v2-tokenizer](https://huggingface.co/bosonai/higgs-audio-v2-tokenizer) | ~200MB | ✅ | Tokenizer model |
 
-## IndexTTS-2
+## Higgs Audio v3
+
+| Component | Source | Size | Auto-Download | Notes |
+|---|---|---|---|---|
+| higgs-audio-v3-tts-4b | [bosonai/higgs-audio-v3-tts-4b](https://huggingface.co/bosonai/higgs-audio-v3-tts-4b) | ~8GB | ✅ | Official 4B multilingual controllable TTS model |
+
+## IndexTTS 2 / 2.5
 
 | Component | Source | Size | Auto-Download | Notes |
 |---|---|---|---|---|
 | IndexTTS-2 | [IndexTeam/IndexTTS-2](https://huggingface.co/IndexTeam/IndexTTS-2) | Multiple files | ✅ | Main TTS engine |
+| IndexTTS-2.5 | [IndexTeam/IndexTTS-2.5](https://huggingface.co/IndexTeam/IndexTTS-2.5) | ~5.49GB | ✅ | Multilingual backend with bundled codec and official feature-duration scaling |
 | w2v-bert-2.0 | [facebook/w2v-bert-2.0](https://huggingface.co/facebook/w2v-bert-2.0) | ~2GB | ✅ | Semantic feature extractor |
 | qwen0.6bemo4-merge | Included with IndexTTS-2 | Included | ✅ | Text emotion model bundle |
 
@@ -90,6 +97,8 @@ Use this as the canonical list of model repositories/links for offline setup.
 | Component | Source | Size | Auto-Download | Notes |
 |---|---|---|---|---|
 | granite-4.0-1b-speech | [ibm-granite/granite-4.0-1b-speech](https://huggingface.co/ibm-granite/granite-4.0-1b-speech) | ~4.6GB | ✅ | Main Granite ASR / AST model |
+| granite-speech-4.1-2b | [ibm-granite/granite-speech-4.1-2b](https://huggingface.co/ibm-granite/granite-speech-4.1-2b) | ~4.6GB | ✅ | Granite Speech 4.1 2B multilingual model (supports Japanese) |
+| granite-speech-4.1-2b-plus | [ibm-granite/granite-speech-4.1-2b-plus](https://huggingface.co/ibm-granite/granite-speech-4.1-2b-plus) | ~4.6GB | ✅ | Plus variant supporting native speaker diarization and word-level timestamps |
 | Qwen3-ForcedAligner-0.6B | [Qwen/Qwen3-ForcedAligner-0.6B](https://huggingface.co/Qwen/Qwen3-ForcedAligner-0.6B) | N/A | ✅ | Optional custom word-level timestamps/SRT path; reused from Qwen folder |
 
 ## Step Audio EditX
@@ -106,15 +115,53 @@ Use this as the canonical list of model repositories/links for offline setup.
 | echo-tts-base (model + PCA state) | [jordand/echo-tts-base](https://huggingface.co/jordand/echo-tts-base) | ~5.3GB | ✅ | pytorch_model.safetensors + pca_state.safetensors |
 | fish-s1-dac-min (audio codec) | [jordand/fish-s1-dac-min](https://huggingface.co/jordand/fish-s1-dac-min) | ~1.8GB | ✅ | pytorch_model.safetensors — audio codec required by Echo-TTS |
 
+## Fish Audio S2 Pro
+
+| Component | Source | Size | Auto-Download | Notes |
+|---|---|---|---|---|
+| S2 Pro | [fishaudio/s2-pro](https://huggingface.co/fishaudio/s2-pro) | ~10.3GB | ✅ | Official 4B model and codec; BNB INT8/NF4 are optional load-time quantization modes that reuse these files; non-commercial license |
+| S2 Pro FP8 | [drbaph/s2-pro-fp8](https://huggingface.co/drbaph/s2-pro-fp8) | ~8.0GB | ✅ | Community weight-only FP8 checkpoint; BF16 activations; RTX 4090/5090-class CUDA GPU required |
+
+## Dots TTS
+
+| Component | Source | Size | Auto-Download | Notes |
+|---|---|---|---|---|
+| dots.tts-base | [rednote-hilab/dots.tts-base](https://huggingface.co/rednote-hilab/dots.tts-base) | ~6GB | ✅ | Official base checkpoint with tokenizer, vocoder, speaker encoder, and latent stats |
+| dots.tts-soar | [rednote-hilab/dots.tts-soar](https://huggingface.co/rednote-hilab/dots.tts-soar) | ~6GB | ✅ | Official SOAR checkpoint for higher-quality zero-shot cloning |
+| dots.tts-mf | [rednote-hilab/dots.tts-mf](https://huggingface.co/rednote-hilab/dots.tts-mf) | ~6GB | ✅ | Official MeanFlow-distilled checkpoint for faster inference |
+
+## DramaBox
+
+| Component | Source | Size | Auto-Download | Notes |
+|---|---|---|---|---|
+| DramaBox DiT + audio components | [ResembleAI/Dramabox](https://huggingface.co/ResembleAI/Dramabox) | ~8.5GB | ✅ | Official merged DramaBox transformer and LTX audio VAE/vocoder components |
+| Gemma 3 12B 4-bit text encoder | [unsloth/gemma-3-12b-it-bnb-4bit](https://huggingface.co/unsloth/gemma-3-12b-it-bnb-4bit) | ~7.8GB | ✅ | Official pre-quantized text encoder; loaded locally with no HF cache fallback |
+
+## OmniVoice
+
+| Component | Source | Size | Auto-Download | Notes |
+|---|---|---|---|---|
+| OmniVoice | [k2-fsa/OmniVoice](https://huggingface.co/k2-fsa/OmniVoice) | ~3.7GB | ✅ | Official main model including bundled audio_tokenizer assets |
+
 ## MOSS-TTS
 
 | Component | Source | Size | Auto-Download | Notes |
 |---|---|---|---|---|
 | MOSS-TTS-Local-Transformer | [OpenMOSS-Team/MOSS-TTS-Local-Transformer](https://huggingface.co/OpenMOSS-Team/MOSS-TTS-Local-Transformer) | ~6.1GB | ✅ | Official 1.7B local-transformer model |
 | MOSS-TTS | [OpenMOSS-Team/MOSS-TTS](https://huggingface.co/OpenMOSS-Team/MOSS-TTS) | ~17GB | ✅ | Official 8B delay model |
+| MOSS-TTS-v1.5 | [OpenMOSS-Team/MOSS-TTS-v1.5](https://huggingface.co/OpenMOSS-Team/MOSS-TTS-v1.5) | ~17GB | ✅ | Current official 8B delay model with 31 languages and more stable voice cloning |
+| MOSS-TTS v1.5 Voice Acting 8B (Community - LAION) | [laion/moss-tts-v1.5-8b-voice-acting](https://huggingface.co/laion/moss-tts-v1.5-8b-voice-acting) | ~17GB | ✅ | Third-party full MOSS-TTS v1.5 fine-tune for expressive voice acting; not an official OpenMOSS model |
+| MOSS-VoiceGenerator | [OpenMOSS-Team/MOSS-VoiceGenerator](https://huggingface.co/OpenMOSS-Team/MOSS-VoiceGenerator) | ~4.2GB | ✅ | Official 1.7B reference-free voice-design model |
 | MOSS-TTSD-v1.0 | [OpenMOSS-Team/MOSS-TTSD-v1.0](https://huggingface.co/OpenMOSS-Team/MOSS-TTSD-v1.0) | ~18GB | ✅ | Official 8B native multi-speaker dialogue model |
+| MOSS-SoundEffect | [OpenMOSS-Team/MOSS-SoundEffect](https://huggingface.co/OpenMOSS-Team/MOSS-SoundEffect) | ~17GB | ✅ | Official MOSS v1 prompt-only sound-effect checkpoint; uses the shared MOSS audio tokenizer |
 | MOSS-Audio-Tokenizer | [OpenMOSS-Team/MOSS-Audio-Tokenizer](https://huggingface.co/OpenMOSS-Team/MOSS-Audio-Tokenizer) | ~8.5GB | ✅ | Shared official codec required by MOSS-TTS |
 
+## MOSS-SoundEffect v2
+
+| Component | Source | Size | Auto-Download | Notes |
+|---|---|---|---|---|
+| MOSS-SoundEffect-v2.0 | [OpenMOSS-Team/MOSS-SoundEffect-v2.0](https://huggingface.co/OpenMOSS-Team/MOSS-SoundEffect-v2.0) | ~11.2GB | ✅ | Official DiT + DAC VAE + Qwen3 text-encoder sound-effect pipeline |
+
 ## RVC
 
 | Component | Source | Size | Auto-Download | Notes |

docs/MODEL_LAYOUTS.md

@@ -203,6 +203,57 @@ Notes:
 
 - Main model, tokenizer assets, and speech stack auto-download.
 
+## Higgs Audio v3
+
+```text
+ComfyUI/models/TTS/higgs_audio_v3/
+└── higgs-audio-v3-tts-4b/
+    ├── config.json
+    ├── tokenizer.json
+    ├── tokenizer_config.json
+    ├── chat_template.jinja
+    ├── model.safetensors
+    ├── model.safetensors.index.json
+    └── LICENSE
+```
+
+Notes:
+
+- Downloads into its own `higgs_audio_v3` folder.
+- Requires the main Transformers 5 environment.
+- Reference transcript `.txt` files are optional but improve cloning quality.
+
+## DramaBox
+
+```text
+ComfyUI/models/TTS/dramabox/
+├── DramaBox/
+    ├── dramabox-dit-v1.safetensors
+    ├── dramabox-audio-components.safetensors
+    ├── assets/
+    │   └── silence_latent_frame.pt
+    └── gemma-3-12b-it-bnb-4bit/
+        ├── config.json
+        ├── model-00001-of-00002.safetensors
+        ├── model-00002-of-00002.safetensors
+        ├── model.safetensors.index.json
+        └── tokenizer and processor files...
+└── loras/
+    └── <adapter_name>/
+        ├── adapter_config.json
+        └── adapter_model.safetensors
+```
+
+Notes:
+
+- Both repositories download directly into the organized suite folder.
+- Transformers is forced into local-only loading after download.
+- Requires NVIDIA CUDA. Fast mode targets approximately 24GB VRAM; experimental staged/sequential modes can run with less.
+- Integrated training exports LoRA adapters into `dramabox/loras/<adapter_name>/`.
+- Training jobs, normalized datasets, preprocessing caches, checkpoints, and logs are stored under `ComfyUI/output/tts_audio_suite_training/dramabox/`.
+- The LTX-2 Community License requires a paid license for entities with at
+  least USD 10 million in annual revenue.
+
 ## CosyVoice3
 
 ```text
@@ -243,7 +294,11 @@ Notes:
 ```text
 ComfyUI/models/TTS/moss_tts/
 ├── MOSS-TTS-Local-Transformer/
+├── MOSS-TTS-v1.5/
+├── moss-tts-v1.5-8b-voice-acting/  # Community - LAION
 ├── MOSS-TTS/
+├── MOSS-VoiceGenerator/
+├── MOSS-SoundEffect/
 ├── MOSS-TTSD-v1.0/
 ├── MOSS-Audio-Tokenizer/
 └── loras/
@@ -256,11 +311,35 @@ Notes:
 
 - `MOSS-Audio-Tokenizer` is required by the official TTS and TTSD variants.
 - `MOSS-TTS-Local-Transformer` is the smaller 1.7B model.
-- `MOSS-TTS` is the official 8B delay model and is much larger.
+- `MOSS-TTS-v1.5` is the current 8B delay model with 31-language support.
+- `moss-tts-v1.5-8b-voice-acting` is an optional third-party LAION full fine-tune for expressive speech, not an official OpenMOSS model.
+- Other compatible full checkpoints placed here are discovered from their `config.json`; unsupported MOSS architectures are rejected explicitly.
+- `MOSS-TTS` is the legacy official 8B delay model.
+- `MOSS-VoiceGenerator` is the 1.7B voice-design provider used by Voice Designer.
+- `MOSS-SoundEffect` is the v1 sound-effect checkpoint used through the MOSS-TTS engine and 🌩️ Sound Effects.
 - `MOSS-TTSD-v1.0` is the official 8B native multi-speaker dialogue model.
 - Integrated training currently exports LoRA adapters into `moss_tts/loras/<adapter_name>/`.
 - Training jobs, temporary manifests, and checkpoints are stored under `ComfyUI/output/tts_audio_suite_training/moss_tts/`.
 
+## MOSS-SoundEffect v2
+
+```text
+ComfyUI/models/TTS/moss_soundeffect_v2/
+└── MOSS-SoundEffect-v2.0/
+    ├── model_index.json
+    ├── scheduler/
+    ├── text_encoder/
+    ├── tokenizer/
+    ├── transformer/
+    └── vae/
+```
+
+Notes:
+
+- This is a separate v2 diffusion family, not a MOSS-TTS checkpoint variant.
+- It runs in the configured ComfyUI environment; the official Apache-2.0 inference package is bundled without modifying its dependencies.
+- The 🌩️ Sound Effects node limits generation to the official 30-second maximum.
+
 ## Granite ASR
 
 ```text
@@ -277,6 +356,7 @@ ComfyUI/models/TTS/granite_asr/
 Notes:
 
 - Granite downloads into its own `granite_asr` folder.
+- `granite-speech-4.1-2b-plus` adds native speaker diarization and native word-level timestamps, but drops Japanese.
 - If Granite word timestamps are enabled, it lazily reuses `Qwen3-ForcedAligner-0.6B` from the Qwen ASR folder instead of duplicating that model.
 
 ## Echo-TTS
@@ -294,3 +374,93 @@ Notes:
 
 - Both components are required and auto-downloaded on first use.
 - License: CC-BY-NC-SA (non-commercial).
+
+## Fish Audio S2 Pro
+
+```text
+ComfyUI/models/TTS/fish_audio_s2_pro/
+├── codec.pth
+├── config.json
+├── model-00001-of-00002.safetensors
+├── model-00002-of-00002.safetensors
+├── model.safetensors.index.json
+└── tokenizer.json
+```
+
+Optional FP8 variant:
+
+```text
+ComfyUI/models/TTS/fish_audio_s2_pro_fp8/
+├── codec.pth
+├── config.json
+├── model.safetensors
+├── quantization_info.json
+└── tokenizer.json
+```
+
+The complete official repository metadata and tokenizer files are downloaded alongside these files. License: Fish Audio Research License (non-commercial without a separate commercial license).
+
+The `s2-pro-bnb-int8` and `s2-pro-bnb-nf4` options reuse `fish_audio_s2_pro/` and quantize its official checkpoint while loading. They do not download another model copy and require `bitsandbytes`.
+
+## Dots TTS
+
+```text
+ComfyUI/models/TTS/dots_tts/
+├── dots.tts-base/
+├── dots.tts-soar/
+└── dots.tts-mf/
+```
+
+Typical checkpoint contents:
+
+```text
+ComfyUI/models/TTS/dots_tts/dots.tts-soar/
+├── added_tokens.json
+├── chat_template.jinja
+├── config.json
+├── latent_stats.pt
+├── llm_config.json
+├── merges.txt
+├── model.safetensors
+├── speaker_encoder.safetensors
+├── special_tokens_map.json
+├── tokenizer.json
+├── tokenizer_config.json
+├── vocab.json
+└── vocoder.safetensors
+```
+
+Notes:
+
+- Downloads into a dedicated `dots_tts/` folder.
+- Native sample rate is 48kHz.
+- Main-environment support works on Transformers 5; on Windows, `normalize_text` falls back to no-op if `WeTextProcessing` is unavailable.
+
+## OmniVoice
+
+```text
+ComfyUI/models/TTS/omnivoice/
+└── OmniVoice/
+```
+
+Main model contents:
+
+```text
+ComfyUI/models/TTS/omnivoice/OmniVoice/
+├── chat_template.jinja
+├── config.json
+├── model.safetensors
+├── tokenizer.json
+├── tokenizer_config.json
+└── audio_tokenizer/
+    ├── config.json
+    ├── model.safetensors
+    └── preprocessor_config.json
+```
+
+Notes:
+
+- Downloads into a dedicated `omnivoice/` folder.
+- Native sample rate is 24kHz.
+- The main OmniVoice model stays in the main Transformers 5 environment.
+- Voice cloning requires reference audio plus explicit reference text in this suite.

docs/MOSS_TRAINING_GUIDE.md

@@ -8,9 +8,14 @@ Use this if `🧾 MOSS Dataset Rows` feels unclear.
 
 Current first training slice supports:
 
-- **MOSS-TTS 8B (Delay)**
+- **MOSS-TTS 8B v1.0 and v1.5 (Delay)**
+- **LAION MOSS-TTS v1.5 Voice Acting 8B community full checkpoint (Delay, compatibility path; training results not yet validated by the suite maintainers)**
 - **LoRA adapter training**
 
+The model selected on the connected MOSS engine is used for dataset preparation and training. Prepare the dataset again after switching between v1.0 and v1.5.
+
+The LAION Voice Acting checkpoint uses the same Delay architecture and can use this LoRA training path, but the suite maintainers have not completed an inference or training run with its full weights. Treat it as community-tested support and report results or incompatibilities.
+
 It does **not** currently support:
 
 - Local 1.7B training
@@ -21,12 +26,29 @@ It does **not** currently support:
 
 Current ComfyUI flow:
 
-1. `🎞️ MOSS Clip Staging`
+1. `🎞️ Training Clip Staging`
 2. `🧾 MOSS Dataset Rows`
 3. `📦 MOSS Dataset Prep`
 4. `🎛️ MOSS Training Config`
 5. `🎓 Model Training`
 
+If clips and transcripts are already prepared on disk, you can skip the first two
+nodes. Set `dataset_source` on `📦 MOSS Dataset Prep` to a folder containing
+same-name audio and text pairs:
+
+```text
+my_dataset/
+├── clip001.wav
+├── clip001.txt
+├── clip002.flac
+└── clip002.txt
+```
+
+Each `.txt` file must contain the transcript spoken in its matching audio file.
+Folder scanning supports WAV, FLAC, MP3, OGG, and M4A. Subfolders are ignored
+unless `recursive_folder_scan` is enabled. Existing JSONL manifest paths continue
+to work unchanged.
+
 ## The Important Fields
 
 ### `text_lines`
@@ -227,7 +249,7 @@ If you do not have a separate validation manifest:
 
 If you want the least confusing starting point:
 
-- use `🎞️ MOSS Clip Staging`
+- use `🎞️ Training Clip Staging`
 - use `🧾 MOSS Dataset Rows`
 - fill only `text_lines`
 - leave `reference_clip_lines` blank

docs/MULTILINE_TTS_TAG_EDITOR_GUIDE.md

@@ -6,7 +6,10 @@ For related topics, also see:
 
 - [Character Switching Guide](CHARACTER_SWITCHING_GUIDE.md)
 - [Parameter Switching Guide](PARAMETER_SWITCHING_GUIDE.md)
-- [Inline Edit Tags User Guide](INLINE_EDIT_TAGS_USER_GUIDE.md)
+- [Step Audio EditX Inline Tags User Guide](INLINE_EDIT_TAGS_USER_GUIDE.md)
+- [Higgs Audio v3 Inline Tags](HIGGS_AUDIO_V3_INLINE_TAGS.md)
+- [CosyVoice3 Tags Guide](COSYVOICE3_TAGS_GUIDE.md)
+- [OmniVoice Native Tags Guide](OMNIVOICE_TAGS_GUIDE.md)
 
 ## What This Editor Is For
 
@@ -17,7 +20,7 @@ It supports:
 - Character switching tags
 - Language switching tags
 - Per-segment parameter overrides
-- Inline Step Audio EditX tags
+- Engine-aware inline tags for Step Audio EditX, Higgs Audio v3, CosyVoice3, and OmniVoice
 - Presets and edit history
 - SRT-aware highlighting and timing editing
 
@@ -48,11 +51,15 @@ Useful behavior:
 - Character names are inserted at the caret or wrapped around the current selection
 - Language and speaker can be combined in one tag
 - Parameters can be stacked with `|`
+- Keep pauses separate and before the tag they precede: `[pause:1s] [Alice|temperature:0.7]`
+- `Format` moves a pause nested with character or parameter parts into that standalone form
 - Presets can store either quick snippets or reusable speaker setups
 
-## Inline Edit Tags
+## Inline Tags
 
-The editor also supports inline Step Audio EditX tags such as:
+The editor has an engine-aware `Inline Tags` panel.
+
+Step Audio EditX examples:
 
 ```text
 <Laughter>
@@ -61,7 +68,36 @@ The editor also supports inline Step Audio EditX tags such as:
 <restore:1@2>
 ```
 
-Use the dedicated inline-edit controls in the sidebar when you do not want to type these by hand.
+Higgs Audio v3 examples:
+
+```text
+<|emotion:amusement|>
+<|style:whispering|>
+<|prosody:pause|>
+```
+
+CosyVoice3 examples:
+
+```text
+<breath>
+<laughing>that was funny</laughing>
+```
+
+OmniVoice examples:
+
+```text
+<laughter>
+<sigh>
+<question-ei>
+```
+
+Use the dedicated inline tag controls in the sidebar when you do not want to type these by hand.
+
+Important differences:
+
+- `Step Audio EditX` tags are post-process controls
+- `Higgs Audio v3`, `CosyVoice3`, and `OmniVoice` tags are native generation controls
+- `OmniVoice` editor insertion uses suite-default angle-tag aliases and the processor converts them internally to official native tags during generation
 
 ## SRT Editing
 

docs/New Engines Guides/03_DECIDE_ENGINE_SCOPE.md

@@ -22,9 +22,14 @@ Produce a short scope document with:
 4. Which non-native suite-added features are proposed, if any?
 5. Which existing engine should be used as implementation reference?
 6. What files/modules will likely need changes?
-7. What manual tests will prove the integration works?
+7. Does this engine stay in the Main Environment, or does it need runtime isolation?
+8. What manual tests will prove the integration works?
 ```
 
+## Runtime Decision Order
+
+Test the official package with `--no-deps` in the main T5 environment first. Prefer a small compatibility patch when practical; use the existing shared T4 SHRED runtime only if T5 is genuinely incompatible. Never create another environment or download/reinstall Torch or Transformers without explicit maintainer approval.
+
 ## Node Types To Choose From
 
 Decide whether the engine needs:
@@ -85,6 +90,10 @@ Manual tests:
 - TTS Text with pause tags
 - SRT generation with interrupt test
 - Clear VRAM then regenerate
+
+Runtime plan:
+- Main Environment only
+- or Shared Runtime / Dedicated Runtime if dependency conflicts are expected
 ```
 
 If the scope is vague, do not code yet.

docs/New Engines Guides/04_IMPLEMENTATION_ORDER_FOR_LLM.md

@@ -43,6 +43,7 @@ Use Qwen3-TTS for:
 - SRT processor structure that reuses the main TTS processor for generation.
 - Interrupt handling.
 - Progress feedback.
+- Console observability parity, such as input-text echo and useful live generation progress.
 - Unified model loading.
 
 Secondary reference:
@@ -69,6 +70,7 @@ Use VibeVoice when the target engine is long-form and needs duration-aware segme
 
 - Prefer model-appropriate duration/minute-based segmentation when required.
 - Do not force short-form `max_chars_per_chunk` behavior if it conflicts with native long-form generation strategy.
+- If the suite already provides chunk splitting / chunk combination controls in the unified node, reuse those. Do not add engine-local chunk-silence or chunk-combination controls unless the official model has a clearly separate native long-form mechanism.
 
 ## Implementation Order
 
@@ -89,13 +91,43 @@ Follow this order:
 13. Add interrupt checks in long loops.
 14. Add progress feedback for long generation.
 15. Update docs/YAML metadata.
-16. Run manual ComfyUI tests.
+16. Run automated and live ComfyUI validation. Use FL-MCP-assisted validation when it is installed and connected; otherwise perform the same checks manually. Follow `tests/FL_MCP_VALIDATION.md`.
 17. Run the required parity checklist.
 
+## Live ComfyUI Validation Rule
+
+Passing imports or pytest is not enough for a new engine. Validate it in the canonical Windows ComfyUI installation after implementation.
+
+If [ComfyUI_FL-MCP](https://github.com/filliptm/ComfyUI_FL-MCP) is available, the LLM should use it to inspect and operate the live ComfyUI instance. Treat it as an optional test driver, not a project dependency and not a substitute for the existing test suite.
+
+The LLM should:
+
+- After changing Python code, restart the canonical Windows ComfyUI process before testing. An already-running process still has the old modules loaded.
+- Use PowerShell to identify the process listening on port `8188`, verify its command line belongs to the canonical ComfyUI `main.py`, stop only that process, and relaunch it with the canonical Windows Python.
+- Wait for `http://127.0.0.1:8188/system_stats` to respond before using FL-MCP.
+- Refresh the existing ComfyUI browser tab after restart and confirm the FL-MCP browser bridge has reconnected before calling canvas-only tools.
+- Confirm the engine node and the relevant unified node are registered.
+- Load or construct the smallest useful workflow.
+- Inspect workflow JSON for the expected node types, links, and widget values.
+- Queue the workflow and wait for completion.
+- Inspect execution history and report the full actionable error if execution fails.
+- Confirm the expected audio output artifact exists.
+- Capture a canvas screenshot for UI and broken-node inspection.
+- Exercise TTS Text and SRT for every TTS engine, plus any other scoped capability.
+- Record what was actually tested, what was skipped, and why.
+
+Screenshots prove only visible workflow state. They do not prove that generation succeeded or that audio is correct. Execution history and output artifacts are required evidence, and the user must still judge subjective audio quality.
+
+Do not install FL-MCP, alter its safety settings, or enable destructive tools unless the user authorizes it. If FL-MCP is unavailable, report that fact and follow the manual fallback in `tests/FL_MCP_VALIDATION.md`.
+
+Repeat the edit, restart, reconnect, and validation cycle after every implementation fix that changes imported Python code. Frontend-only changes may require a hard browser refresh as well. Do not claim that a fix was tested against a ComfyUI process started before the fix was written.
+
 ## Architecture Rule
 
 Unified nodes should stay thin.
 
+Reference engines are examples only. Every engine must have dedicated processors and adapters; share only engine-neutral utilities.
+
 Do not put hundreds of lines of engine-specific orchestration into:
 
 - `nodes/unified/tts_text_node.py`
@@ -126,6 +158,15 @@ The SRT processor should not:
 - Reimplement character switching, pause tags, parameter switching, cache, or reference-audio handling differently from the TTS processor.
 - Add separate fake SRT-only parameters that do not exist in the normal TTS path or the suite timing layer.
 
+## Chunking Rule
+
+If an engine needs long-text chunking, prefer the suite chunking and shared chunk combiner unless the official model exposes a real native long-form/chunking system that is materially different.
+
+Do not:
+
+- Add duplicate engine-node controls for silence between chunks, chunk combination method, or similar suite-owned chunk-assembly behavior.
+- Bypass the shared chunk combiner with a custom engine-local join path unless there is a documented official reason.
+
 ## Registration Areas To Check
 
 The LLM should inspect and update the relevant registration files for the chosen scope:

docs/New Engines Guides/05_REQUIRED_PARITY_CHECKLIST.md

@@ -25,6 +25,9 @@ Downloads and dependencies:
 - Do models download into organized ComfyUI/models/TTS/ folders?
 - Did you prevent silent downloads into random cache folders?
 - Did you document dependency conflicts or install.py changes?
+- Did you test `--no-deps` on Main/T5 and simple patching before falling back to the shared T4 runtime?
+- Did you avoid creating another environment or downloading/reinstalling Torch or Transformers without explicit maintainer approval?
+- If runtime isolation is needed, did you document the default mode and the reason in YAML/README?
 
 Audio format:
 - Did you verify the model native sample rate?
@@ -44,16 +47,19 @@ Text features:
 - Did you wire parameter switching for real native parameters?
 - Did you avoid adding fake unsupported parameters?
 - Did you remove misleading controls like language selection if unsupported?
+- If the engine supports language control, did you normalize the user-facing language UI/logs honestly and map them back to official backend codes without breaking segment/alias language switching?
+- Did you reuse the suite chunking/chunk-combination controls instead of inventing duplicate engine-local chunk UI or chunk-silence parameters?
 
 SRT features:
 - If this is a TTS engine, did you implement the SRT path?
 - Did you add interrupt checks in long SRT loops?
-- Does SRT use the shared timing/assembly systems correctly?
+- Does SRT use the shared timing/assembly/reporting systems correctly, using the real existing APIs instead of guessed helper names?
 - Did you avoid unnecessary timing-mode restrictions?
 - Does character switching work in SRT if text character tags are supported?
 
 UX:
 - Did you add progress feedback for long generation?
+- Did you preserve useful console observability parity instead of making the engine run silently (for example input-text echo and live it/s/progress)?
 - Are error messages clear when required inputs are missing?
 - Are tooltips honest about limitations, license, VRAM, and expected text length?
 
@@ -62,6 +68,7 @@ Docs and metadata:
 - Did you regenerate docs/tables if metadata changed?
 - Did you document model paths and download sources?
 - Did you add examples or notes for special native features?
+- If the engine needs a secondary environment, does the engine table clearly show that?
 
 Manual tests:
 - Did TTS Text work?
@@ -71,8 +78,23 @@ Manual tests:
 - Did character tags work?
 - Did pause tags work?
 - Did parameter switching work?
-- Did Clear VRAM then regenerate work?
+- Did Clear VRAM then regenerate work, and did unload actually tear down runtime/cache state instead of only moving weights to CPU?
 - Did interrupt/cancel work in long generation?
+
+Live ComfyUI integration evidence:
+- Was validation run in the canonical Windows ComfyUI environment?
+- Was ComfyUI restarted after the final Python changes, and can you show that the tested process started after those edits?
+- After restart, was the browser refreshed and the FL-MCP browser bridge confirmed connected before canvas checks?
+- If FL-MCP was available, did you follow tests/FL_MCP_VALIDATION.md and identify which MCP checks were run?
+- If FL-MCP was unavailable, did you perform and document the equivalent manual checks instead of claiming MCP validation?
+- Are the engine node and relevant unified nodes present without import or registration errors?
+- Does the saved workflow JSON contain the expected node types, links, and widget values?
+- Does the workflow load without missing/broken node state?
+- Was the workflow queued, and does execution history show successful completion or the full actionable error?
+- Does each successful run produce the expected audio output artifact?
+- Was a canvas screenshot captured for UI inspection without treating the screenshot as execution proof?
+- Did you record tested, failed, and skipped cases, including the reason for every skip?
+- Did a human assess subjective audio quality separately?
 ```
 
 ## Important Failures This Prevents

docs/New Engines Guides/NEW_ENGINE_IMPLEMENTATION_GUIDE.md

@@ -308,11 +308,13 @@ from utils.downloads.unified_downloader import UnifiedDownloader
 
 **File:** `engines/adapters/[engine_name]_adapter.py`
 
-#### Step 4: Create Engine Configuration Node
-
-**File:** `nodes/engines/[engine_name]_engine_node.py`
-
-### Phase 2: Unified Systems Integration
+#### Step 4: Create Engine Configuration Node
+
+**File:** `nodes/engines/[engine_name]_engine_node.py`
+
+**Model dropdown rule:** Always keep canonical/downloadable model choices visible and add detected installations as separate `local:ModelName` choices; selecting a canonical choice should reuse its organized local installation when available, not replace or hide either choice.
+
+### Phase 2: Unified Systems Integration
 
 #### Step 5: Integrate with Unified Model Loading
 
@@ -462,11 +464,12 @@ Also to test, requirements and dependencies need to be added.
 - [ ] Character switching works with `[CharacterName] text`
 - [ ] Language switching works (if applicable)
 - [ ] Pause tags work with `[pause:1.5s]`
-- [ ] Caching works (same input = cached output)
-- [ ] Model auto-download works
-- [ ] VRAM management works (model unloads)
-- [ ] Different parameter combinations work
-- [ ] **Interrupt handling works** - User can stop SRT generation and it stops within ~1 segment
+- [ ] Caching works (same input = cached output)
+- [ ] Model auto-download works
+- [ ] VRAM management works (model unloads)
+- [ ] Different parameter combinations work
+- [ ] Engine prints a standard `Settings:` summary with the active generation/load parameters
+- [ ] **Interrupt handling works** - User can stop SRT generation and it stops within ~1 segment
 
 ### Phase 4: SRT Implementation
 
@@ -645,4 +648,4 @@ Update the engines comparison table.
 - Model management logic
 - Audio format conversion utilities
 
----
+---

docs/New Engines Guides/README.md

@@ -67,6 +67,7 @@ A new engine is not done until the answer to these questions is yes:
 - Does it use the suite architecture instead of stuffing engine logic into unified nodes?
 - Does it load models through the unified model interface?
 - Does model unload/reload/Clear VRAM work?
+- Did someone explicitly decide whether this engine belongs in the Main Environment or needs runtime isolation?
 - Does it support generated audio cache where applicable?
 - Does it handle ComfyUI audio shape and sample rate correctly?
 - Does it wire character tags, pause tags, and parameter switching where applicable?

docs/New Engines Guides/fails_to_avoid_TTS_Engine_Implementation.md

@@ -56,11 +56,22 @@
 - **Cache key pattern**: `audio_cache.generate_cache_key(engine_type, text=..., audio_component=..., **all_params)`
 - **Duration calculation**: Update `_calculate_duration()` with engine sample rate (e.g., 24000 for Step Audio EditX, F5-TTS)
 
+### Engine Settings Logging
+- **Missing standard print**: New engines should print the usual `Settings:` summary once per run with the active generation/load parameters so validation can confirm what actually executed
+- **Resolved prompt preview**: Reuse `utils.voice.character_logging.format_resolved_character_block()` for boxed text previews so logs show the voice that will actually speak, not only the parser alias
+
 ### Model Lifecycle - __del__ Destructor
 - **CRITICAL**: Remove `__del__` from engine classes - causes automatic unload after generation ends (when object goes out of scope)
 - **Pattern**: F5-TTS and ChatterBox don't have `__del__`, only IndexTTS and StepAudio did (wrong)
 - **Correct behavior**: Models stay in VRAM for reuse; only unload on explicit Clear VRAM button, engine switch, or parameter change
 
+### Clear VRAM - Runtime Teardown vs CPU Offload
+- **Wrong**: Assuming every engine is safe to unload with only `.to('cpu')`
+- **Symptom**: Clear VRAM frees only part of VRAM; most memory disappears only after closing ComfyUI
+- **Cause**: Runtime-heavy engines can keep compiled graphs, prompt caches, workspaces, or runtime objects alive after CPU offload
+- **Correct**: Add an engine-specific unload handler that clears internal caches, drops runtime/model references, then runs `gc.collect()`, `torch.cuda.empty_cache()`, and `torch.cuda.ipc_collect()` when available
+- **Prevention**: If the engine uses warmup, `torch.compile`, streaming caches, or a custom runtime object, explicitly test Clear VRAM and reload instead of assuming generic offload parity
+
 ### Generation Info Reports - Modular Pattern
 - **CRITICAL - Consistency required**: ALL engines must use `ChunkCombiner` + `ChunkTimingHelper` for generation reports
 - **Correct pattern** (from ChatterBox 23-Lang, F5-TTS):
@@ -74,10 +85,27 @@
 ### Segment Parameter Registration
 - **Forgot to add**: New engines to `PARAMETER_ENGINES` in `utils/text/segment_parameters.py` for `[seed:X]` tag support
 
+### Language UI Normalization
+- **Wrong**: Exposing raw backend language codes in the engine node just because the official runtime accepts them
+- **Symptom**: New engine UI looks inconsistent with the suite, logs show backend codes instead of readable names, and character-tag / alias languages do not map cleanly to node config values
+- **Correct**: Keep user-facing language controls normalized to honest display names when the suite already does that, then map them back to the official backend codes in the adapter/runtime layer
+- **Prevention**: Reuse or add a small engine-local language helper so node UI, processor logs, segment language overrides, and adapter normalization all agree on the same mapping
+
 ### SRT Processor Import Pattern
 - **Wrong**: Using `from nodes.xxx import` - fails with `'nodes' is not a package`
 - **Correct**: Use `importlib.util.spec_from_file_location()` for cross-module imports in dynamically loaded processors
 
+### SRT Shared API Guessing
+- **Wrong**: Inventing shared helper names like `AudioAssemblyEngine.assemble(...)`, `TimingReportGenerator`, or `AdjustedSRTGenerator`
+- **Correct**: Open a working SRT processor first and match the exact API that already exists in `utils/timing/assembly.py`, `utils/timing/engine.py`, and `utils/timing/reporting.py`
+- **Real pattern**:
+  - `stretch_to_fit` → `TimedAudioAssembler.assemble_timed_audio(...)`
+  - `pad_with_silence` → `AudioAssemblyEngine(...).assemble_with_overlaps(...)`
+  - `concatenate` → `TimingEngine.calculate_concatenation_adjustments(...)` + `assemble_concatenation(...)`
+  - `smart_natural` → `TimingEngine.calculate_smart_timing_adjustments(...)` + `assemble_smart_natural(...)`
+  - reporting / adjusted SRT → `SRTReportGenerator().generate_timing_report(...)` and `generate_adjusted_srt_string(...)`
+- **Prevention**: Never assume shared helper names. Grep existing SRT processors and copy the real call pattern exactly.
+
 ### ComfyUI Audio Dict Format
 - **Wrong**: Passing `audio_tensor` directly to functions expecting raw tensor
 - **Correct**: Extract waveform with `audio_tensor.get('waveform')` - ComfyUI audio is `{"waveform": tensor, "sample_rate": int}`

docs/OMNIVOICE_TAGS_GUIDE.md

@@ -0,0 +1,73 @@
+# OmniVoice Native Tags Guide
+
+OmniVoice has its own official inline square-bracket control tokens upstream, but **this suite does not expose `[]` for OmniVoice tags**.
+
+Inside TTS Audio Suite, you can write the suite-default angle-tag aliases in text:
+
+```text
+<laughter>
+<sigh>
+<question-ei>
+```
+
+The OmniVoice processor converts those aliases internally to the official OmniVoice native form before generation:
+
+```text
+[laughter]
+[sigh]
+[question-ei]
+```
+
+Do **not** type OmniVoice non-verbal tags in `[]` form in suite text. In this suite, `[]` belongs to character, language, parameter, and pause syntax.
+
+## Supported Non-Verbal Tags
+
+Official OmniVoice non-verbal meanings exposed by this suite through `<>` aliases:
+
+```text
+laughter
+sigh
+confirmation-en
+question-en
+question-ah
+question-oh
+question-ei
+question-yi
+surprise-ah
+surprise-oh
+surprise-wa
+surprise-yo
+dissatisfaction-hnn
+```
+
+Examples:
+
+```text
+[Alice] <laughter> You really got me there.
+[Bob] <sigh> Fine, let's try again.
+[Narrator] <question-ei> Really?
+[Narrator] <surprise-oh> I didn't expect that.
+```
+
+## Important Behavior
+
+- OmniVoice uses native generation tags here, not Step Audio EditX inline post-processing.
+- User-facing suite syntax stays in `<>` form for OmniVoice non-verbal tags.
+- `[]` is reserved for suite structural syntax like `[Alice]`, `[en:Alice]`, `[pause:1s]`, and parameter switching.
+- Do not rely on Step-style `<Laughter:2>` or `<emotion:happy>` semantics in the OmniVoice text path.
+- If you want Step Audio EditX as a second pass on OmniVoice output, use the separate `🎨 Audio Editor` node manually after generation.
+
+## Multiline Tag Editor
+
+The `🏷️ Multiline TTS Tag Editor` has a dedicated `OmniVoice` mode in the `Inline Tags` panel.
+
+- The editor inserts suite-default angle-tag aliases like `<laughter>`
+- The processor converts them internally to official OmniVoice square tags during generation
+- The editor does not encourage raw OmniVoice `[]` input because `[]` is suite syntax
+
+## Sources
+
+This behavior follows the official OmniVoice documentation for non-verbal symbols and pronunciation control:
+
+- [OmniVoice GitHub README](https://github.com/k2-fsa/OmniVoice)
+- [OmniVoice Hugging Face model card](https://huggingface.co/k2-fsa/OmniVoice)

docs/PARAMETER_SWITCHING_GUIDE.md

@@ -95,6 +95,20 @@ Parameters are applied **only to the current segment** and automatically revert
 | `sound_event` | — | string | text | Whole-segment sound event hint |
 | `ambient_sound` | — | string | text | Whole-segment ambient sound hint |
 
+#### MOSS Sound Effects
+
+MOSS-SoundEffect v1 uses the applicable MOSS-TTS parameters above. Both sound-effect engines also support a duration override:
+
+| Parameter | Alias | Engines | Type | Range | Description |
+|-----------|-------|---------|------|-------|-------------|
+| `duration_seconds` | `seconds` | v1, v2 | float | 0.5-300 | Duration of the sound segment |
+| `inference_steps` | `steps` | v2 | int | 1-150 | Diffusion steps |
+| `cfg` | — | v2 | float | 0.0-20.0 | Prompt guidance strength |
+| `sigma_shift` | — | v2 | float | 0.0-10.0 | Flow-matching schedule shift |
+| `negative_prompt` | `negative`, `neg` | v2 | string | text | Sounds or qualities to discourage |
+
+See the [Sound Effects Guide](SOUND_EFFECTS_GUIDE.md) for pauses, crossfades, long-duration chunking, and complete examples.
+
 #### ChatterBox & ChatterBox Official 23-Lang
 | Parameter | Alias | Type | Range | Description |
 |-----------|-------|------|-------|-------------|
@@ -121,13 +135,67 @@ Parameters are applied **only to the current segment** and automatically revert
 | `top_p` | `topp` | float | 0.0-1.0 | Nucleus sampling probability |
 | `inference_steps` | `steps` | int | 1-100 | Number of inference steps |
 
-#### IndexTTS-2
+#### IndexTTS 2 / 2.5
 | Parameter | Alias | Type | Range | Description |
 |-----------|-------|------|-------|-------------|
 | `cfg` | — | float | 0.0-20.0 | CFG strength |
 | `top_p` | `topp` | float | 0.0-1.0 | Nucleus sampling probability |
 | `top_k` | `topk` | int | 1-100 | Top-k sampling |
-| `emotion_alpha` | — | float | 0.0-1.0 | Emotion control strength |
+| `emotion_alpha` | — | float | 0.0-1.0 | Shared audio/vector/text emotion intensity |
+| `duration_factor` | `dur_factor` | float | 0.5-2.0 | Official IndexTTS-2.5 internal feature-duration scaling; 0.5 shorter/faster, 2.0 longer/slower |
+
+`duration_factor` is a 2.5-only upstream parameter. It uses nearest-neighbor scaling inside the semantic length regulator after speech codes are generated. It is not natural prosody planning, exact-seconds targeting, waveform playback-speed control, or an inference-performance control. IndexTTS continues to use the suite's ordinary final timing modes in TTS SRT.
+
+Switching the engine node between IndexTTS-2 and IndexTTS-2.5 invalidates the cached Text/SRT processor and model identity. `language`, `duration_factor`, and `text_normalization` also participate in the generated-audio cache identity, so changing a supported 2.5 generation parameter cannot return audio produced with the previous setting.
+
+IndexTTS-2 also supports inline emotion controls. Named unsigned values replace
+that dimension; explicitly signed values adjust the connected vector:
+
+```text
+[sad:0.7|calm:0.2] Absolute values for this segment.
+[sad:+0.3|calm:-0.2] Adjust the connected vector for this segment.
+```
+
+All eight values can be supplied in the official order `happy, angry, sad,
+afraid, disgusted, melancholic, surprised, calm`:
+
+```text
+[vector:0,0,0.7,0,0,0.4,0,0.2] Absolute replacement.
+[vector:+0,+0,+0.3,+0,+0,+0,+0,-0.2] Relative adjustment.
+```
+
+Full relative vectors require an explicit sign on every value. Results are
+clamped to IndexTTS-2's supported range and revert to the connected vector at
+the next segment.
+
+Text emotion can use a saved preset or quoted text. `{seg}` is expanded with
+the current segment before QwenEmotion analysis:
+
+```text
+[emotion:restrained_anger] A saved preset.
+[emotion:"Quiet grief masking frustration"] A direct description.
+[emotion:"Infer nervous anticipation from this line: {seg}"] Dynamic analysis.
+```
+
+Click a numeric emotion tag in the TTS Tag Editor to open a contextual radar
+directly beside that tag. The editor also creates and manages text
+presets in `models/TTS/IndexTTS/emotion_presets.json`.
+
+IndexTTS-2 has separate engine inputs for these sources: connect vector or text
+emotion to `emotion_control` and audio emotion references to `emotion_audio`.
+Both may be connected simultaneously; IndexTTS-2 blends them during emotion
+conditioning. Inline vector/text controls override the connected vector/text
+values for their segment, while `[Character:emotion_ref]` selects a
+segment-local audio reference that can still blend with vector/text emotion.
+When an emotion control is inserted with the caret inside a character/audio tag,
+the editor appends or updates it as another pipe parameter, for example
+`[Bob:br_ivan_raiva3|sad:+0.25]`.
+
+The tag editor's quick-swap palette is engine-aware: parameter choices are
+filtered to the selected inline engine, while each engine's supported native
+emotion/style/prosody/sound tags use their own replacement choices. Named
+emotion presets can be swapped from the text; quoted `[emotion:"..."]` text is
+intentionally left as direct editable content rather than treated as a preset.
 
 ---
 
@@ -158,6 +226,20 @@ Important:
 - These are whole-segment controls
 - They are not positional inline effects
 - Keep `<>` free for true inline post-processing tags like Step Audio EditX
+
+### DramaBox Prompt Templates
+
+`prompt_template` (alias `template`) applies a `{seg}` wrapper and enables
+templating for that segment automatically:
+
+```text
+[Narrator|template:A woman whispers, "{seg}"] This line is whispered.
+[Narrator] This line returns to the DramaBox engine-node settings.
+```
+
+The template should include `{seg}`. If it is omitted, DramaBox warns once and
+appends `"{seg}"` automatically. A separate inline enable parameter is not
+required.
 
 ### Per-Segment Fine-Tuning in SRT
 

docs/SOUND_EFFECTS_GUIDE.md

@@ -0,0 +1,113 @@
+# 🌩️ Sound Effects Guide
+
+The `🌩️ Sound Effects` node generates non-speech audio from a written description. It works with any connected engine that advertises sound-effect support.
+
+## Engines
+
+| Model | Engine node | Notes |
+|---|---|---|
+| MOSS-SoundEffect v1 | `⚙️ MOSS-TTS Engine` | Autoregressive MOSS 8B model |
+| MOSS-SoundEffect v2 | `⚙️ MOSS SoundEffect v2 Engine` | 48 kHz diffusion model; up to 30 seconds per native generation |
+
+Connecting a speech-only engine stops with a user-facing compatibility error.
+
+## Basic workflow
+
+1. Select a sound-effect model on its engine node.
+2. Connect the engine to `🌩️ Sound Effects`.
+3. Describe the sound rather than words to be spoken.
+4. Choose the duration and seed, then queue the workflow.
+
+Example:
+
+```text
+Heavy rain hitting a metal rooftop, distant rolling thunder, occasional wind gusts.
+```
+
+Descriptions generally work best when they state the source, environment, distance, texture, and progression of the sound.
+
+## Timeline segments
+
+Separate descriptions with parameter tags to generate multiple segments and concatenate them:
+
+```text
+[seconds:4|seed:42] Bright application startup chime. [seconds:2|cfg:5] Low, dark shutdown tone.
+```
+
+`duration_seconds` is the default duration for every generated segment. `[seconds:X]` overrides it for the following segment.
+
+Newlines also create segments, as they do in TTS. They are optional because a parameter tag can start another segment on the same line.
+
+## Pauses
+
+Use a standalone pause tag to insert exact silence:
+
+```text
+[seconds:4] Startup chime. [pause:1.2] [seconds:2] Shutdown tone.
+```
+
+The aliases `[wait:X]` and `[stop:X]` are also accepted. Durations may use seconds or milliseconds:
+
+```text
+[wait:500ms]
+```
+
+Keep pauses separate from parameter tags:
+
+```text
+[pause:1.2] [cfg:7.5] Thunder crack.
+```
+
+Do not combine them as `[pause:1.2|cfg:7.5]`.
+
+## Crossfade and long sounds
+
+`crossfade_seconds` overlaps adjacent generated segments to soften their join. Set it to `0` for a hard join.
+
+A pause creates an exact silent boundary, so crossfade is not applied across that pause.
+
+MOSS-SoundEffect v2 has a native 30-second generation limit. Longer requested segments are generated as overlapping chunks, joined with the selected crossfade, and trimmed to the requested duration.
+
+## Per-segment parameters
+
+Common parameters:
+
+| Tag | Engines | Purpose |
+|---|---|---|
+| `seed` | v1, v2 | Generated variation |
+| `seconds` / `duration_seconds` | v1, v2 | Segment duration |
+| `temperature` | v1 | Sampling randomness |
+| `top_p`, `top_k` | v1 | Sampling limits |
+| `repetition_penalty` | v1 | Discourage repetition |
+| `duration_tokens` | v1 | Native duration-token control |
+| `max_new_tokens` | v1 | Generation token limit |
+| `steps` / `inference_steps` | v2 | Diffusion steps |
+| `cfg` | v2 | Prompt guidance strength |
+| `sigma_shift` | v2 | Flow-matching schedule shift |
+| `negative_prompt`, `negative`, `neg` | v2 | Sounds or qualities to discourage |
+
+Example using a negative prompt:
+
+```text
+[seconds:8|cfg:5|neg:speech, music] Dense forest ambience with insects and distant birds.
+```
+
+Unsupported parameters are ignored with a warning rather than being sent blindly to the engine.
+
+## Seed and cache behavior
+
+- `seed: 0` chooses a random seed.
+- Reusing a positive seed with identical settings makes the request repeatable.
+- With audio caching enabled, an identical segment and configuration can reuse its generated audio.
+- Changing the description, seed, duration, engine configuration, or inline parameters invalidates that cached result.
+
+## MOSS-SoundEffect v2 first-run compilation
+
+The v2 DiT uses `torch.compile`. Its first generation may spend several minutes compiling before progress begins. Compatible compilation artifacts are cached and can be reused across later ComfyUI sessions.
+
+This compile delay is separate from model downloading and normal generation time.
+
+## Related guides
+
+- [Per-Segment Parameter Switching](PARAMETER_SWITCHING_GUIDE.md)
+- [Multiline TTS Tag Editor](MULTILINE_TTS_TAG_EDITOR_GUIDE.md)

docs/TEXT_CHUNKING_GUIDE.md

@@ -0,0 +1,47 @@
+# Text Chunking Guide
+
+TTS Audio Suite supports long text generation through the shared unified chunking system.
+
+## What It Does
+
+- splits long text into smaller generation chunks
+- preserves sentence boundaries when possible
+- can fall back to comma-based splitting for very long sentences
+- combines generated audio using the suite chunk combiner
+
+## Main Controls
+
+- `enable_chunking`
+  - turns chunking on or off
+- `max_chars_per_chunk`
+  - character limit per chunk
+- `chunk_combination_method`
+  - `auto`
+  - `concatenate`
+  - `silence_padding`
+  - `crossfade`
+- `silence_between_chunks_ms`
+  - silence inserted when the selected combination method uses silence padding
+
+## Auto Combination Behavior
+
+When `chunk_combination_method=auto`, the suite picks a join strategy based on text size:
+
+- longer text tends toward `silence_padding`
+- medium text tends toward `crossfade`
+- shorter chunked text tends toward `concatenate`
+
+The exact choice is made by the shared chunk combiner, not by engine-local custom logic.
+
+## Important Architecture Rule
+
+Chunking is a suite feature.
+
+- new engines should reuse the shared chunking and chunk-combination controls
+- engines should not invent duplicate local chunk-silence or chunk-join settings unless the official model has a real distinct native long-form system
+
+## Related Docs
+
+- [Example Workflows](../README.md#-example-workflows)
+- [Character Switching Guide](./CHARACTER_SWITCHING_GUIDE.md)
+- [Multiline TTS Tag Editor Guide](./MULTILINE_TTS_TAG_EDITOR_GUIDE.md)

docs/Upstream_Engine_Tracking/INDEXTTS2_UPDATE_LOG.md

@@ -11,6 +11,33 @@ This document tracks updates applied to our bundled IndexTTS-2 code from the ups
 
 ---
 
+## 2026-08-11: IndexTTS-2.5 Version Integration
+
+**Official sources:** `index-tts/index-tts` commit `b5ea881bec284b72f0b1cc04e0a724ff0c6b93e9`; model snapshot `ba2480d9f7f629eb18f6acaebb357679d9ba88a4`
+
+### Changes applied
+
+- Added IndexTTS-2.5 as a selectable version of the existing `index_tts` engine.
+- Bundled the official 25 Hz semantic codec, multilingual tokenizer, Japanese G2P, and NeMo normalization bridge.
+- Preserved suite dual-source audio plus vector/text emotion blending.
+- Added Chinese, English, Japanese, Spanish, and Arabic conditioning.
+- Added the official 2.5-only `duration_factor`, documented honestly as nearest-neighbor internal semantic-feature scaling rather than natural prosody or exact-duration planning.
+- Deliberately excluded IndexTTS-2.5 from TTS SRT's native-duration option; the suite-owned exact-seconds extrapolation was removed after source and listening review.
+- Kept legacy IndexTTS-2 checkpoints, FP16 loading, MaskGCT, workflows, and node identity intact.
+- Pinned the audited Hugging Face model revision and retained the main Transformers 5 environment.
+- Added model-aware Text/SRT processor and audio-cache identities so switching 2.0/2.5 or a 2.5 generation parameter cannot reuse stale output.
+- Documented the suite's manual finding that 2.5 is not a universal cloning-quality upgrade: 2.0 may retain speaker resemblance better under strong different-speaker emotion transfer.
+
+### Validation status
+
+- [x] Python compilation
+- [x] Bundled backend import under `TTS_SUITE_TEST_VENV_PYTHON`
+- [x] Full checkpoint download and live ComfyUI generation
+- [x] Manual audio-quality review of the official duration factor and 2.0/2.5 speaker resemblance
+- [x] Live 2.5 → 2.0 model switching after processor-cache invalidation fix
+
+---
+
 ## 2025-09-18: Major Update - Cache & Emotion Improvements
 
 **Reference commit range:** `8336824..64cb31a` (September 11 → September 18, 2025)
@@ -207,6 +234,47 @@ This document tracks updates applied to our bundled IndexTTS-2 code from the ups
 
 ---
 
+---
+
+## 2026-07-11: Upstream Audit Before Release
+
+**Upstream repository checked:** `index-tts/index-tts` (`main`)
+**Upstream head observed:** `b5bd657` (2026-07-08)
+**Check performed:** 2026-07-11
+
+### Relevant upstream changes reviewed
+
+| Commit | Upstream change | Bundled status | Release action |
+|---|---|---|---|
+| `843972e` | Coerce QwenEmotion JSON emotion scores to `float` and reject non-numeric values clearly | **Applied** in bundled `clamp_score()` | Run a text-emotion generation with numeric-string classifier output |
+| `b154a1b` | WebUI text/vector preset save/load management | **Already covered differently** by the ComfyUI-native preset manager and `emotion_presets.json` integration | No direct port needed |
+| `b5bd657` | Expose `--accel` and `--torch-compile`, add optional acceleration extras and WebUI settings | **Applied selectively**; suite now forwards `use_accel` into the bundled GPT path, while retaining ComfyUI-owned dependency handling | Validate acceleration fallback on compatible and non-accelerated setups |
+| `7264ce2` | Improve IndexTTS-2 model resource checks and HF cache handling | **Suite-owned downloader differs** and needs a separate comparison if download failures are reported | No blind copy into bundled code |
+
+### Findings
+
+- No upstream change was found that invalidates the current eight-emotion vector order,
+  Qwen text-emotion syntax, audio-reference blending, or the suite's inline tag format.
+- The upstream QwenEmotion string-score fix is directly relevant to the suite's text-emotion
+  path and should be applied before a release.
+- The upstream acceleration work is not a drop-in replacement because this repository
+  bundles and adapts IndexTTS-2. The suite already contains the acceleration modules, but
+  `utils/models/unified_model_interface.py` should be checked so `use_accel` reaches the
+  bundled `IndexTTS2` constructor.
+- Upstream WebUI presets are not copied verbatim: the suite's ComfyUI editor has a richer
+  vector/radar, inline-tag, sidebar, and filesystem preset implementation.
+
+### Release follow-up checklist
+
+- [x] Apply the upstream `clamp_score()` numeric coercion.
+- [x] Pass `use_accel` through the unified IndexTTS-2 factory; validate fallback behavior.
+- [ ] Run a Qwen text-emotion generation using numeric-string classifier output.
+- [ ] Verify acceleration flags on a compatible CUDA installation and on a setup without
+      optional acceleration dependencies.
+- [ ] Recheck bundled model-resource validation against the current upstream `check` logic.
+
+---
+
 ## Next Update Check: 2025-11-20
 
 **Monitoring:** Watch for commits to `indextts/infer_v2.py` in https://github.com/index-tts/index-tts
@@ -222,4 +290,4 @@ git log --oneline 1d5d079..HEAD -- indextts/infer_v2.py
 - Additional performance optimizations
 - Bug fixes in new acceleration code
 - Breaking API changes
-- Model loading improvements
+- Model loading improvements

docs/qwen3_tts_optimizations.md

@@ -4,24 +4,32 @@ This document explains how to enable torch.compile optimizations for Qwen3-TTS t
 
 ## Requirements
 
-**CRITICAL**: torch.compile optimizations require specific versions tested with the reference implementation:
-
-- **PyTorch 2.10.0+** with CUDA 13.0 support
-- **triton-windows 3.6.0+** (Windows) or **triton 3.6.0+** (Linux)
-- Python 3.12 (recommended)
+**CRITICAL**: torch.compile optimizations require specific versions tested with the reference implementation:
+
+- **PyTorch 2.10.0+** with CUDA 13.0 support
+- **triton-windows 3.6.0+** (Windows) or **triton 3.6.0+** (Linux)
+- **Visual Studio C++ Build Tools** on Windows (a real `cl.exe` toolchain, not just stale PATH entries)
+- Python 3.12 (recommended)
 
 **Note**: PyTorch 2.8.x and 2.9.x have compatibility issues with triton on Windows. Only PyTorch 2.10+ is supported.
 
 ### Installation
 
-#### Windows (Tested Configuration)
-```bash
-# Upgrade PyTorch to 2.10 with CUDA 13.0
-pip install --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130
-
-# Install triton-windows 3.6.x
-pip install -U "triton-windows<3.7"
-```
+#### Windows (Tested Configuration)
+```bash
+# Upgrade PyTorch to 2.10 with CUDA 13.0
+pip install --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130
+
+# Install triton-windows 3.6.x
+pip install -U "triton-windows<3.7"
+```
+
+Also install **Visual Studio Build Tools** with the C++ toolchain:
+
+- Workload: `Desktop development with C++`
+- Or at minimum the MSVC x64/x86 build tools component
+
+If `cl.exe` is missing, `torch.compile` in the isolated Qwen runtime will fall back to the non-compiled path.
 
 **Verification:**
 ```bash
@@ -46,7 +54,12 @@ python -c "import triton; print(triton.__version__)"
 
 ## Environment Configuration
 
-**No special configuration needed** - PyTorch 2.10 works with the default cudaMallocAsync allocator when using `compile_mode="default"`.
+**Main environment**: no special configuration beyond a working compiler toolchain.
+
+**Shared / Dedicated Runtime isolation**:
+- TTS Audio Suite now tries to detect Visual Studio automatically and inject a proper MSVC build environment into the isolated worker
+- This is meant to preserve `torch.compile` parity with old direct Transformers 4 environments
+- It still depends on the Build Tools actually being installed
 
 **Note**: The `reduce-overhead` mode (which includes automatic CUDA graphs) may still have issues with cudaMallocAsync. Use `compile_mode="default"` for best compatibility.
 
@@ -88,10 +101,20 @@ Once the environment is configured:
 pip install --upgrade torchvision --index-url https://download.pytorch.org/whl/cu130
 ```
 
-### Error: "cudaMallocAsync does not support checkPoolLiveAllocations"
-**Cause**: Using `compile_mode="reduce-overhead"` which enables automatic CUDA graphs.
-
-**Solution**: Use `compile_mode="default"` instead (recommended for Windows).
+### Error: "cudaMallocAsync does not support checkPoolLiveAllocations"
+**Cause**: Using `compile_mode="reduce-overhead"` which enables automatic CUDA graphs.
+
+**Solution**: Use `compile_mode="default"` instead (recommended for Windows).
+
+### Error: "Failed to find C compiler. Please specify via CC environment variable."
+**Cause**: Triton/Inductor could not find a usable Windows C++ toolchain.
+
+**Solution**:
+1. Install or repair Visual Studio Build Tools with the C++ workload
+2. Restart ComfyUI so the new environment is visible
+3. If using Shared/Dedicated Runtime, let TTS Audio Suite auto-detect the toolchain
+
+If detection still fails, your machine state is broken, not just the Python package stack.
 
 ### Warning: "TensorFloat32 tensor cores... not enabled"
 **Info**: This is just a performance hint, not an error. TF32 is automatically enabled by the extension for better performance.
@@ -110,12 +133,13 @@ pip install --upgrade torchvision --index-url https://download.pytorch.org/whl/c
 
 ### Key Findings
 
-- **torch.compile alone provides ~1.7x speedup** with mode="default"
-- **Manual CUDA graphs add almost nothing** on top of torch.compile
-- **reduce-overhead and max-autotune fail on Windows** (work on Linux)
-- **First generation is slower** (~30-60s compilation overhead)
-- **Subsequent generations are fast** (compiled kernels cached)
-- **Model unloading works correctly** even with CUDA graphs enabled
+- **torch.compile alone provides ~1.7x speedup** with mode="default"
+- **Manual CUDA graphs add almost nothing** on top of torch.compile
+- **reduce-overhead and max-autotune fail on Windows** (work on Linux)
+- **First generation is slower** (~30-60s compilation overhead)
+- **Subsequent generations are fast** (compiled kernels cached)
+- **Model unloading works correctly** even with CUDA graphs enabled
+- **Shared Runtime should preserve these gains** if the Windows compiler toolchain is installed and detected correctly
 
 ## Notes
 

engines/adapters/__init__.py

@@ -40,6 +40,42 @@ except Exception as e:
         def __init__(self, *args, **kwargs):
             raise ImportError(f"Echo-TTS adapter not available: {e}")
 
+try:
+    from .dots_tts_adapter import DotsTTSEngineAdapter
+    DOTS_TTS_ADAPTER_AVAILABLE = True
+except Exception as e:
+    DOTS_TTS_ADAPTER_AVAILABLE = False
+    class DotsTTSEngineAdapter:
+        def __init__(self, *args, **kwargs):
+            raise ImportError(f"Dots TTS adapter not available: {e}")
+
+try:
+    from .dramabox_adapter import DramaBoxEngineAdapter
+    DRAMABOX_ADAPTER_AVAILABLE = True
+except Exception as e:
+    DRAMABOX_ADAPTER_AVAILABLE = False
+    class DramaBoxEngineAdapter:
+        def __init__(self, *args, **kwargs):
+            raise ImportError(f"DramaBox adapter not available: {e}")
+
+try:
+    from .fish_audio_s2_adapter import FishAudioS2Adapter
+    FISH_AUDIO_S2_ADAPTER_AVAILABLE = True
+except Exception as e:
+    FISH_AUDIO_S2_ADAPTER_AVAILABLE = False
+    class FishAudioS2Adapter:
+        def __init__(self, *args, **kwargs):
+            raise ImportError(f"Fish Audio S2 adapter not available: {e}")
+
+try:
+    from .omnivoice_adapter import OmniVoiceEngineAdapter
+    OMNIVOICE_ADAPTER_AVAILABLE = True
+except Exception as e:
+    OMNIVOICE_ADAPTER_AVAILABLE = False
+    class OmniVoiceEngineAdapter:
+        def __init__(self, *args, **kwargs):
+            raise ImportError(f"OmniVoice adapter not available: {e}")
+
 try:
     from .moss_tts_adapter import MossTTSEngineAdapter
     MOSS_TTS_ADAPTER_AVAILABLE = True
@@ -49,9 +85,24 @@ except Exception as e:
         def __init__(self, *args, **kwargs):
             raise ImportError(f"MOSS-TTS adapter not available: {e}")
 
+try:
+    from .higgs_audio_v3_adapter import HiggsAudioV3EngineAdapter
+    HIGGS_AUDIO_V3_ADAPTER_AVAILABLE = True
+except Exception as e:
+    HIGGS_AUDIO_V3_ADAPTER_AVAILABLE = False
+    class HiggsAudioV3EngineAdapter:
+        def __init__(self, *args, **kwargs):
+            raise ImportError(f"Higgs Audio v3 adapter not available: {e}")
+
 __all__ = [
     'ChatterBoxEngineAdapter', 'F5TTSEngineAdapter', 'CosyVoiceAdapter', 'EchoTTSEngineAdapter',
-    'MossTTSEngineAdapter',
+    'DotsTTSEngineAdapter', 'DramaBoxEngineAdapter', 'OmniVoiceEngineAdapter',
+    'MossTTSEngineAdapter', 'HiggsAudioV3EngineAdapter',
     'CHATTERBOX_ADAPTER_AVAILABLE', 'F5TTS_ADAPTER_AVAILABLE', 'COSYVOICE_ADAPTER_AVAILABLE',
-    'ECHO_TTS_ADAPTER_AVAILABLE', 'MOSS_TTS_ADAPTER_AVAILABLE'
+    'ECHO_TTS_ADAPTER_AVAILABLE', 'DOTS_TTS_ADAPTER_AVAILABLE',
+    'DRAMABOX_ADAPTER_AVAILABLE', 'OMNIVOICE_ADAPTER_AVAILABLE',
+    'MOSS_TTS_ADAPTER_AVAILABLE', 'HIGGS_AUDIO_V3_ADAPTER_AVAILABLE',
+    'MossSoundEffectV2Adapter'
 ]
+
+from .moss_soundeffect_v2_adapter import MossSoundEffectV2Adapter

engines/adapters/asr_audio_cpp_adapter.py

@@ -0,0 +1,486 @@
+"""audio.cpp adapter for the Suite's unified ASR pipeline."""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+import time
+from typing import Any, Dict, Iterable, Mapping, Optional
+
+import torch
+
+from utils.asr.types import ASRRequest, ASRResult, ASRSegment, ASRWord
+from utils.audio.processing import AudioProcessingUtils
+
+
+_NATIVE_CHUNK_FAMILIES = {
+    "fun_asr_nano",
+    "higgs_audio_stt",
+    "hviske_asr",
+    "qwen3_asr",
+    "vibevoice_asr",
+    "voxtral_realtime",
+}
+
+# audio.cpp release-0.5.1 keeps stale offline decoder state for these loaders:
+# the first request transcribes normally and later requests return empty text.
+# A fresh owned process is currently the only reliable reset contract.
+_RESTART_BETWEEN_CHUNKS_FAMILIES = {"nemotron_asr", "voxtral_realtime"}
+
+
+def _session(config: Mapping[str, Any]):
+    from utils.audio_cpp.session import get_audio_cpp_session
+
+    return get_audio_cpp_session(dict(config))
+
+
+def _advanced_options(config: Mapping[str, Any]) -> Dict[str, Any]:
+    value = config.get("advanced_options", config.get("request_options", {}))
+    if value in (None, ""):
+        return {}
+    if isinstance(value, str):
+        try:
+            value = json.loads(value)
+        except json.JSONDecodeError as exc:
+            raise ValueError(f"Invalid audio.cpp advanced JSON: {exc.msg}") from exc
+    if not isinstance(value, Mapping):
+        raise ValueError("audio.cpp advanced options must be a JSON object")
+    return dict(value)
+
+
+def _audio_path(audio: Mapping[str, Any]) -> str:
+    waveform = audio.get("waveform")
+    sample_rate = audio.get("sample_rate")
+    if not torch.is_tensor(waveform):
+        raise TypeError("audio.cpp ASR input must contain a waveform tensor")
+    if sample_rate is None or int(sample_rate) <= 0:
+        raise ValueError("audio.cpp ASR input must contain a positive sample_rate")
+    return os.path.abspath(
+        AudioProcessingUtils.save_audio_to_temp_file(waveform, int(sample_rate))
+    )
+
+
+def _waveform_3d(audio: Mapping[str, Any]) -> tuple[torch.Tensor, int]:
+    waveform = audio.get("waveform")
+    sample_rate = int(audio.get("sample_rate") or 0)
+    if not torch.is_tensor(waveform):
+        raise TypeError("audio.cpp ASR input must contain a waveform tensor")
+    if sample_rate <= 0:
+        raise ValueError("audio.cpp ASR input must contain a positive sample_rate")
+    if waveform.ndim == 1:
+        waveform = waveform.unsqueeze(0).unsqueeze(0)
+    elif waveform.ndim == 2:
+        waveform = waveform.unsqueeze(0)
+    elif waveform.ndim != 3:
+        raise ValueError(
+            "audio.cpp ASR waveform must have [samples], [channels, samples], or "
+            "[batch, channels, samples] shape"
+        )
+    if waveform.shape[0] != 1:
+        raise ValueError("audio.cpp ASR accepts one audio item at a time")
+    if waveform.shape[-1] <= 0:
+        raise ValueError("audio.cpp ASR input audio is empty")
+    return waveform.detach().cpu(), sample_rate
+
+
+def _chunk_ranges(
+    total_samples: int,
+    sample_rate: int,
+    chunk_size: int,
+    overlap: int,
+) -> list[tuple[int, int]]:
+    if chunk_size <= 0:
+        return [(0, total_samples)]
+    if overlap < 0:
+        raise ValueError("ASR overlap must be zero or greater")
+    if overlap >= chunk_size:
+        raise ValueError("ASR overlap must be smaller than chunk_size")
+
+    chunk_samples = chunk_size * sample_rate
+    if total_samples <= chunk_samples:
+        return [(0, total_samples)]
+    step_samples = (chunk_size - overlap) * sample_rate
+    ranges = []
+    start = 0
+    while start < total_samples:
+        end = min(start + chunk_samples, total_samples)
+        ranges.append((start, end))
+        if end >= total_samples:
+            break
+        start += step_samples
+    return ranges
+
+
+def _normalized_token(value: str) -> str:
+    return re.sub(r"[^\w]+", "", value, flags=re.UNICODE).casefold()
+
+
+def _merge_transcript(parts: Iterable[str]) -> str:
+    merged: list[str] = []
+    for part in parts:
+        incoming = str(part or "").strip().split()
+        if not incoming:
+            continue
+        if not merged:
+            merged.extend(incoming)
+            continue
+        limit = min(len(merged), len(incoming), 80)
+        duplicate_count = 0
+        for size in range(limit, 0, -1):
+            left = [_normalized_token(token) for token in merged[-size:]]
+            right = [_normalized_token(token) for token in incoming[:size]]
+            if all(left) and left == right:
+                duplicate_count = size
+                break
+        merged.extend(incoming[duplicate_count:])
+    return " ".join(merged).strip()
+
+
+def _offset_words(
+    words: Iterable[ASRWord], offset: float, unique_after: Optional[float]
+) -> list[ASRWord]:
+    shifted = []
+    for word in words:
+        item = ASRWord(start=word.start + offset, end=word.end + offset, text=word.text)
+        if unique_after is not None and (item.start + item.end) / 2.0 < unique_after:
+            continue
+        shifted.append(item)
+    return shifted
+
+
+def _offset_segments(
+    segments: Iterable[ASRSegment], offset: float, unique_after: Optional[float]
+) -> list[ASRSegment]:
+    shifted = []
+    for segment in segments:
+        item = ASRSegment(
+            start=segment.start + offset,
+            end=segment.end + offset,
+            text=segment.text,
+            speaker=segment.speaker,
+        )
+        if unique_after is not None and (item.start + item.end) / 2.0 < unique_after:
+            continue
+        shifted.append(item)
+    return shifted
+
+
+def _seconds(value: Any, sample_rate: int) -> float:
+    try:
+        return max(0.0, float(value) / float(sample_rate))
+    except (TypeError, ValueError, ZeroDivisionError):
+        return 0.0
+
+
+def _words(payload: Mapping[str, Any], sample_rate: int) -> list[ASRWord]:
+    words = []
+    for item in payload.get("words") or []:
+        if not isinstance(item, Mapping):
+            continue
+        text = str(item.get("word", item.get("text", ""))).strip()
+        if not text:
+            continue
+        words.append(
+            ASRWord(
+                start=_seconds(item.get("start_sample"), sample_rate),
+                end=_seconds(item.get("end_sample"), sample_rate),
+                text=text,
+            )
+        )
+    return words
+
+
+def _plain_segments(payload: Mapping[str, Any], sample_rate: int) -> list[ASRSegment]:
+    segments = []
+    for item in payload.get("segments") or []:
+        if not isinstance(item, Mapping):
+            continue
+        text = str(item.get("text", "")).strip()
+        segments.append(
+            ASRSegment(
+                start=_seconds(item.get("start_sample"), sample_rate),
+                end=_seconds(item.get("end_sample"), sample_rate),
+                text=text,
+            )
+        )
+    return segments
+
+
+def _speaker_segments(payload: Mapping[str, Any], sample_rate: int) -> list[ASRSegment]:
+    segments = []
+    for item in payload.get("speaker_turns") or []:
+        if not isinstance(item, Mapping):
+            continue
+        speaker = str(item.get("speaker_id", "")).strip()
+        if speaker and not speaker.lower().startswith("speaker"):
+            speaker = f"Speaker {speaker}"
+        segments.append(
+            ASRSegment(
+                start=_seconds(item.get("start_sample"), sample_rate),
+                end=_seconds(item.get("end_sample"), sample_rate),
+                text=str(item.get("text", "")).strip(),
+                speaker=speaker or None,
+            )
+        )
+    return segments
+
+
+def _attach_words(segments: Iterable[ASRSegment], words: Iterable[ASRWord]) -> None:
+    segment_list = list(segments)
+    for word in words:
+        midpoint = (word.start + word.end) / 2.0
+        target = next(
+            (segment for segment in segment_list if segment.start <= midpoint <= segment.end),
+            None,
+        )
+        if target is not None:
+            target.words.append(word)
+
+
+class AudioCppASREngineAdapter:
+    """Normalize audio.cpp transcript/timing output into ``ASRResult``."""
+
+    def __init__(self, engine_data: Dict[str, Any]):
+        self.engine_data = dict(engine_data)
+        self.config = dict(engine_data.get("config", engine_data))
+
+    def _session_config(self) -> Dict[str, Any]:
+        config = dict(self.config)
+        if str(config.get("connection_mode", "auto")).lower() != "external_server":
+            config["requested_task"] = "asr"
+            config["task"] = "asr"
+        return config
+
+    def transcribe(self, req: ASRRequest) -> ASRResult:
+        if req.task != "transcribe":
+            raise ValueError(
+                "audio.cpp release-0.5.1 ASR loaders support transcription, not the "
+                "Unified ASR translate mode"
+            )
+
+        config = self._session_config()
+        family = str(config.get("family", "")).strip()
+        warnings: list[str] = []
+        notes: list[str] = []
+        options = _advanced_options(config)
+
+        # VibeVoice-ASR owns diarization across its full recording. Independent
+        # Suite requests can restart speaker numbering, so preserve its native
+        # chunking only for this mode. All other ASR uses Suite-side windows.
+        native_diarization = (
+            family == "vibevoice_asr" and req.diarization and req.chunk_size > 0
+        )
+        if native_diarization:
+            options.setdefault("audio_chunk_mode", "fixed")
+            options.setdefault("audio_chunk_seconds", int(req.chunk_size))
+            if req.overlap > 0:
+                notes.append(
+                    "VibeVoice-ASR diarization uses native chunking to preserve speaker "
+                    "identity; the Suite overlap setting is not applied."
+                )
+        elif family in _NATIVE_CHUNK_FAMILIES:
+            options.setdefault("audio_chunk_mode", "none")
+
+        if req.timestamps == "word" and family == "qwen3_asr":
+            session_options = config.get("session_options") or {}
+            aligner = session_options.get("qwen3_asr.forced_aligner_model_path")
+            if aligner:
+                options["return_timestamps"] = True
+            else:
+                warnings.append(
+                    "Qwen3-ASR word timestamps require the optional Qwen3 Forced Aligner; "
+                    "transcription continued without downloading that auxiliary model."
+                )
+
+        waveform, source_rate = _waveform_3d(req.audio)
+        ranges = (
+            [(0, waveform.shape[-1])]
+            if native_diarization
+            else _chunk_ranges(
+                waveform.shape[-1], source_rate, int(req.chunk_size), int(req.overlap)
+            )
+        )
+        session = _session(config)
+        if str(getattr(session, "task", "asr")) != "asr":
+            raise ValueError(
+                f"audio.cpp model '{session.model_id}' is configured for task "
+                f"'{session.task}', not ASR"
+            )
+        restart_between_chunks = (
+            len(ranges) > 1 and family in _RESTART_BETWEEN_CHUNKS_FAMILIES
+        )
+        if restart_between_chunks and not bool(getattr(session, "owned", False)):
+            raise RuntimeError(
+                f"audio.cpp release-0.5.1 {family} returns empty text after its first "
+                "offline request. Suite-side chunking therefore requires a managed "
+                "audio.cpp server so the Suite can reset it between chunks. Set "
+                "connection_mode to managed, or set ASR chunk_size to 0 when using "
+                "an external server."
+            )
+        if restart_between_chunks:
+            notes.append(
+                f"audio.cpp release-0.5.1 {family} requires a managed server reset "
+                "between Suite chunks to avoid empty repeated-request results."
+            )
+
+        display_family = family or "external model"
+        print(f"🎧 audio.cpp ASR: Transcribing with {display_family}...")
+        if len(ranges) > 1:
+            notes.append(
+                f"Suite-side ASR chunking used {len(ranges)} windows of "
+                f"{int(req.chunk_size)}s with {int(req.overlap)}s overlap."
+            )
+            print(
+                f"🧩 audio.cpp ASR: {len(ranges)} chunks "
+                f"({int(req.chunk_size)}s, {int(req.overlap)}s overlap)"
+            )
+
+        payloads: list[Mapping[str, Any]] = []
+        chunk_timings: list[Mapping[str, Any]] = []
+        chunk_diagnostics: list[Dict[str, Any]] = []
+        started_at = time.time()
+        for index, (start, end) in enumerate(ranges, start=1):
+            if index > 1 and restart_between_chunks:
+                print(
+                    f"🔄 audio.cpp ASR: Resetting {family} session for chunk "
+                    f"{index}/{len(ranges)}"
+                )
+                session.restart_owned_runtime()
+            chunk_waveform = waveform[..., start:end]
+            chunk_rms = float(torch.sqrt(torch.mean(chunk_waveform.float().square())).item())
+            chunk_peak = float(chunk_waveform.float().abs().max().item())
+            temp_path = _audio_path({
+                "waveform": chunk_waveform,
+                "sample_rate": source_rate,
+            })
+            try:
+                request: Dict[str, Any] = {"audio": temp_path, "options": dict(options)}
+                if req.language:
+                    request["language"] = req.language
+                result = session.run(request)
+                payload = result.raw if isinstance(result.raw, Mapping) else {}
+                payloads.append(payload)
+                if isinstance(payload.get("timing"), Mapping):
+                    chunk_timings.append(payload["timing"])
+                chunk_diagnostics.append({
+                    "index": index,
+                    "start": round(start / source_rate, 3),
+                    "end": round(end / source_rate, 3),
+                    "rms": round(chunk_rms, 6),
+                    "peak": round(chunk_peak, 6),
+                    "text": str(payload.get("text", "")).strip(),
+                    "characters": len(str(payload.get("text", "")).strip()),
+                    "upstream_timing": (
+                        dict(payload["timing"])
+                        if isinstance(payload.get("timing"), Mapping)
+                        else None
+                    ),
+                })
+            finally:
+                try:
+                    os.remove(temp_path)
+                except FileNotFoundError:
+                    pass
+            if len(ranges) > 1:
+                chunk_chars = len(str(payload.get("text", "")).strip())
+                print(
+                    f"   ASR chunk {index}/{len(ranges)} complete "
+                    f"({chunk_chars} chars, RMS {chunk_rms:.4f}, peak {chunk_peak:.4f})"
+                )
+
+        words: list[ASRWord] = []
+        speaker_segments: list[ASRSegment] = []
+        plain_segments: list[ASRSegment] = []
+        overlap_seconds = float(req.overlap) if len(ranges) > 1 else 0.0
+        for index, ((start, _end), payload) in enumerate(zip(ranges, payloads)):
+            offset = start / source_rate
+            unique_after = offset + overlap_seconds if index > 0 else None
+            words.extend(_offset_words(_words(payload, source_rate), offset, unique_after))
+            speaker_segments.extend(
+                _offset_segments(
+                    _speaker_segments(payload, source_rate), offset, unique_after
+                )
+            )
+            plain_segments.extend(
+                _offset_segments(_plain_segments(payload, source_rate), offset, unique_after)
+            )
+
+        if req.diarization:
+            segments = speaker_segments
+            if segments:
+                _attach_words(segments, words)
+            else:
+                warnings.append(
+                    f"audio.cpp {family or 'ASR model'} returned no speaker-attributed turns."
+                )
+                segments = plain_segments
+        elif req.timestamps == "word" and words:
+            segments = [
+                ASRSegment(start=word.start, end=word.end, text=word.text, words=[word])
+                for word in words
+            ]
+        elif req.timestamps == "word":
+            segments = plain_segments
+        else:
+            segments = []
+
+        text = _merge_transcript(payload.get("text", "") for payload in payloads)
+        if req.diarization and speaker_segments:
+            text = " ".join(
+                f"[{segment.speaker}] {segment.text}" if segment.speaker else segment.text
+                for segment in speaker_segments
+                if segment.text
+            ).strip()
+        if not text and speaker_segments:
+            text = " ".join(segment.text for segment in speaker_segments if segment.text).strip()
+        if req.timestamps == "word" and not words:
+            warnings.append(f"audio.cpp {family or 'ASR model'} returned no word timestamps.")
+        empty_chunks = sum(
+            1 for payload in payloads if not str(payload.get("text", "")).strip()
+        )
+        if len(payloads) > 1 and empty_chunks:
+            warnings.append(
+                f"audio.cpp {family or 'ASR model'} returned no text for "
+                f"{empty_chunks} of {len(payloads)} Suite chunks."
+            )
+
+        raw: Dict[str, Any] = {}
+        if warnings:
+            raw["warnings"] = warnings
+        if notes:
+            raw["notes"] = notes
+        if len(payloads) == 1 and chunk_timings:
+            raw["timing"] = dict(chunk_timings[0])
+        elif len(payloads) > 1:
+            raw["timing"] = {
+                "wall_ms": round((time.time() - started_at) * 1000.0, 3),
+                "suite_chunks": len(payloads),
+                "suite_chunk_size_seconds": int(req.chunk_size),
+                "suite_overlap_seconds": int(req.overlap),
+                "upstream_wall_ms": round(
+                    sum(float(item.get("wall_ms", 0.0)) for item in chunk_timings), 3
+                ),
+            }
+            raw["chunks"] = chunk_diagnostics
+        output_language = next(
+            (
+                str(payload.get("language", "")).strip()
+                for payload in payloads
+                if str(payload.get("language", "")).strip()
+            ),
+            str(req.language or "").strip(),
+        ) or None
+        print(
+            f"✅ audio.cpp ASR: Complete ({len(text)} chars, "
+            f"{len(segments)} timed/speaker segments)"
+        )
+        return ASRResult(
+            text=text,
+            language=output_language,
+            segments=segments,
+            raw=raw or None,
+        )
+
+
+__all__ = ["AudioCppASREngineAdapter"]

engines/adapters/asr_granite_adapter.py

@@ -10,11 +10,211 @@ import torch.nn.functional as F
 import comfy.model_management as model_management
 import comfy.utils
 import time
+import re
 
 from engines.granite_asr.prompting import DEFAULT_TRANSLATE_INSTRUCTION_TEMPLATE
 from utils.asr.types import ASRRequest, ASRResult, ASRSegment, ASRWord
 
 
+def parse_granite_native_timestamps(ts_text: str, chunk_offset: float) -> Tuple[str, List[ASRSegment]]:
+    """
+    Parses native timestamped text generated by ibm-granite/granite-speech-4.1-2b-plus.
+    Reconstructs absolute timestamps by unwrapping modulo-1000 rollovers.
+    """
+    if not ts_text or not ts_text.strip():
+        return "", []
+
+    ts_words = re.split(r"\[T:([^\]]+)\]", ts_text)
+    last_word_end_time = 0.0
+    offset_time = 0.0
+
+    segments: List[ASRSegment] = []
+    plain_words: List[str] = []
+
+    for raw_word, ts_str in zip(ts_words[::2], ts_words[1::2]):
+        word_text = raw_word.strip()
+        if not ts_str:
+            continue
+
+        try:
+            word_end_time = float(ts_str) / 100.0
+        except ValueError:
+            continue
+
+        while word_end_time + offset_time < last_word_end_time:
+            offset_time += 10.0
+        absolute_end_time = word_end_time + offset_time
+        
+        absolute_start_time = last_word_end_time
+        last_word_end_time = absolute_end_time
+
+        if word_text == "_":
+            continue
+
+        if word_text:
+            plain_words.append(word_text)
+            word = ASRWord(
+                start=absolute_start_time + chunk_offset,
+                end=absolute_end_time + chunk_offset,
+                text=word_text,
+            )
+            segments.append(
+                ASRSegment(
+                    start=word.start,
+                    end=word.end,
+                    text=word.text,
+                    words=[word],
+                )
+            )
+
+    plain_text = " ".join(plain_words)
+    return plain_text, segments
+
+
+def parse_granite_diarization(text: str, chunk_offset: float, chunk_duration: float) -> Tuple[str, List[ASRSegment]]:
+    """
+    Parses speaker-attributed ASR text generated by ibm-granite/granite-speech-4.1-2b-plus.
+    Returns:
+        plain_text: Cleaned text without speaker tags.
+        segments: List of ASRSegment instances with speaker property populated.
+    """
+    if not text or not text.strip():
+        return "", []
+
+    # Find speaker tags like [Speaker 1]:, [Speaker 2]:, etc.
+    # Keep the delimiter matching group so we can reconstruct who is speaking.
+    pattern = r"\[(Speaker \d+)\]:?"
+    parts = re.split(pattern, text)
+    
+    segments: List[ASRSegment] = []
+    
+    # Handle text before the first speaker tag, if any
+    first_text = parts[0].strip()
+    if first_text:
+        segments.append(
+            ASRSegment(
+                start=chunk_offset,
+                end=chunk_offset + chunk_duration,
+                text=first_text,
+                speaker=None,
+                words=[],
+            )
+        )
+        
+    for speaker, content in zip(parts[1::2], parts[2::2]):
+        segment_text = content.strip()
+        if not segment_text:
+            continue
+        segments.append(
+            ASRSegment(
+                start=chunk_offset,
+                end=chunk_offset + chunk_duration,
+                text=segment_text,
+                speaker=speaker,
+                words=[],
+            )
+        )
+        
+    plain_text = " ".join(seg.text for seg in segments if seg.text).strip()
+    plain_text = re.sub(r"\s+", " ", plain_text).strip()
+    
+    # Distribute the chunk_duration proportionally across segments based on character length
+    total_len = sum(len(seg.text) for seg in segments if seg.text)
+    if total_len > 0:
+        current_time = chunk_offset
+        for seg in segments:
+            seg_len = len(seg.text) if seg.text else 0
+            duration = chunk_duration * (seg_len / total_len)
+            seg.start = current_time
+            seg.end = current_time + duration
+            current_time += duration
+
+    return plain_text, segments
+
+
+def format_granite_diarization_segments(segments: List[ASRSegment]) -> str:
+    """
+    Normalize Granite speaker attribution to the suite's canonical bracket format:
+    [Speaker 1] text
+    """
+    parts: List[str] = []
+    for seg in segments:
+        text = (seg.text or "").strip()
+        if not text:
+            continue
+        if seg.speaker:
+            parts.append(f"[{seg.speaker}] {text}")
+        else:
+            parts.append(text)
+    return " ".join(parts).strip()
+
+
+def remap_words_to_speaker_segments(speaker_segments: List[ASRSegment], aligned_words: List[ASRWord]) -> None:
+    """
+    Distributes aligned words sequentially to speaker_segments based on text matching.
+    """
+    if not speaker_segments or not aligned_words:
+        return
+
+    word_idx = 0
+    num_words = len(aligned_words)
+
+    for segment in speaker_segments:
+        # Tokenize segment text to find expected words
+        expected_tokens = [w.lower() for w in re.findall(r"\b\w+\b", segment.text) if w]
+        if not expected_tokens:
+            continue
+        
+        segment_words = []
+        match_count = 0
+        limit = len(expected_tokens)
+        
+        while word_idx < num_words and match_count < limit:
+            aligned_word = aligned_words[word_idx]
+            aligned_token_clean = "".join(ch for ch in aligned_word.text.lower() if ch.isalnum())
+            expected_token_clean = "".join(ch for ch in expected_tokens[match_count] if ch.isalnum())
+            
+            if not aligned_token_clean:
+                segment_words.append(aligned_word)
+                word_idx += 1
+                continue
+                
+            if (aligned_token_clean == expected_token_clean or 
+                    expected_token_clean in aligned_token_clean or 
+                    aligned_token_clean in expected_token_clean):
+                segment_words.append(aligned_word)
+                word_idx += 1
+                match_count += 1
+            elif (match_count + 1 < limit and 
+                  ("".join(ch for ch in expected_tokens[match_count + 1] if ch.isalnum()) in aligned_token_clean or
+                   aligned_token_clean in "".join(ch for ch in expected_tokens[match_count + 1] if ch.isalnum()))):
+                # Missed an expected word in alignment, skip the expected word and match current aligned to next expected
+                match_count += 2
+                segment_words.append(aligned_word)
+                word_idx += 1
+            else:
+                # Mismatch, consume to avoid getting stuck
+                segment_words.append(aligned_word)
+                word_idx += 1
+                match_count += 1
+
+        if segment_words:
+            segment.words = segment_words
+            segment.start = segment_words[0].start
+            segment.end = segment_words[-1].end
+            
+    # If there are leftover words, append them to the last segment that had some words
+    if word_idx < num_words and speaker_segments:
+        last_seg = speaker_segments[-1]
+        if last_seg.words:
+            last_seg.words.extend(aligned_words[word_idx:])
+            last_seg.end = aligned_words[-1].end
+        else:
+            last_seg.words = aligned_words[word_idx:]
+            last_seg.start = aligned_words[word_idx].start
+            last_seg.end = aligned_words[-1].end
+
+
 class GraniteASREngineAdapter:
     GRANITE_LANGUAGES = {
         "english": "English",
@@ -48,7 +248,11 @@ class GraniteASREngineAdapter:
     def _normalize_language(self, language: Optional[str]) -> Optional[str]:
         if not language or str(language).lower() == "auto":
             return None
-        return self.GRANITE_LANGUAGES.get(str(language).strip().lower())
+        normalized = self.GRANITE_LANGUAGES.get(str(language).strip().lower())
+        model_name = self.config.get("model_name", "")
+        if normalized == "Japanese" and "plus" in str(model_name).lower():
+            return None
+        return normalized
 
     def _get_model(self):
         from utils.models.unified_model_interface import unified_model_interface, ModelLoadConfig
@@ -63,8 +267,8 @@ class GraniteASREngineAdapter:
         config = ModelLoadConfig(
             engine_name="granite_asr",
             model_type="asr",
-            model_name=self.config.get("model_name", "granite-4.0-1b-speech"),
-            model_path=self.config.get("model_name", "granite-4.0-1b-speech"),
+            model_name=self.config.get("model_name", "granite-speech-4.1-2b"),
+            model_path=self.config.get("model_name", "granite-speech-4.1-2b"),
             device=device,
             additional_params={
                 "precision": self.config.get("dtype", "auto"),
@@ -75,6 +279,7 @@ class GraniteASREngineAdapter:
 
     def _get_forced_aligner(self):
         from utils.models.unified_model_interface import unified_model_interface, ModelLoadConfig
+        from utils.models.factory_config import RUNTIME_MODE_SHARED
         from utils.device import resolve_torch_device
 
         device = self.config.get("device", "auto")
@@ -89,6 +294,8 @@ class GraniteASREngineAdapter:
             model_name="Qwen3-ForcedAligner-0.6B",
             model_path="Qwen3-ForcedAligner-0.6B",
             device=device,
+            runtime_mode=self.config.get("forced_aligner_runtime_mode", RUNTIME_MODE_SHARED),
+            runtime_profile=self.config.get("forced_aligner_runtime_profile"),
             additional_params={
                 "precision": self.config.get("dtype", "auto"),
                 "attn_implementation": self.config.get("attn_implementation", "auto"),
@@ -108,13 +315,12 @@ class GraniteASREngineAdapter:
             waveform = waveform[0]
 
         if sample_rate != 16000:
-            w = waveform.view(1, 1, -1)
-            w = F.interpolate(w, size=int(w.shape[-1] * 16000 / sample_rate), mode="linear", align_corners=False)
-            waveform = w.view(-1)
+            import torchaudio.functional as TA_F
+            waveform = TA_F.resample(waveform, orig_freq=sample_rate, new_freq=16000)
         return waveform
 
     def _alignment_requested(self, req: ASRRequest) -> bool:
-        return req.timestamps == "word" or req.use_forced_aligner
+        return req.timestamps == "word"
 
     def _can_use_aligner(self, req: ASRRequest, normalized_language: Optional[str], warnings: List[str]) -> bool:
         if not self._alignment_requested(req):
@@ -216,6 +422,32 @@ class GraniteASREngineAdapter:
             warnings.append(f"Granite timestamps skipped: Qwen forced aligner failed ({e}).")
             return []
 
+    def _resolve_max_new_tokens(
+        self,
+        req: ASRRequest,
+        *,
+        chunk_num_samples: int,
+        sample_rate: int,
+        timestamps: bool = False,
+        diarization: bool = False,
+    ) -> int:
+        base_max_new_tokens = int(self.config.get("max_new_tokens", req.max_new_tokens or 200))
+        chunk_seconds = max(1.0, float(chunk_num_samples) / float(sample_rate))
+
+        # Native timestamp mode expands output heavily because each spoken word is
+        # followed by a [T:NNN] tag. Speaker attribution also adds extra label tokens.
+        # Keep the user-configured cap as the floor, but auto-raise low budgets so
+        # longer chunks do not get truncated mid-transcript.
+        if timestamps:
+            recommended_floor = int(max(256, min(2048, round(chunk_seconds * 16.0))))
+            return max(base_max_new_tokens, recommended_floor)
+
+        if diarization:
+            recommended_floor = int(max(256, min(2048, round(chunk_seconds * 10.0))))
+            return max(base_max_new_tokens, recommended_floor)
+
+        return base_max_new_tokens
+
     def _run_chunk(
         self,
         runtime,
@@ -224,9 +456,17 @@ class GraniteASREngineAdapter:
         normalized_language: Optional[str],
         target_language: Optional[str],
         translate_instruction_override: Optional[str],
+        timestamps: bool = False,
+        diarization: bool = False,
     ) -> Tuple[str, Optional[str]]:
         generation_kwargs = {
-            "max_new_tokens": int(self.config.get("max_new_tokens", req.max_new_tokens or 200)),
+            "max_new_tokens": self._resolve_max_new_tokens(
+                req,
+                chunk_num_samples=len(chunk_np),
+                sample_rate=16000,
+                timestamps=timestamps,
+                diarization=diarization,
+            ),
             "do_sample": bool(self.config.get("do_sample", False)),
             "num_beams": int(self.config.get("num_beams", 1)),
             "temperature": float(self.config.get("temperature", 1.0)),
@@ -244,6 +484,8 @@ class GraniteASREngineAdapter:
             target_language=target_language,
             translate_instruction_override=translate_instruction_override,
             generation_kwargs=generation_kwargs,
+            timestamps=timestamps,
+            diarization=diarization,
         )
         return result.get("text", ""), result.get("language")
 
@@ -258,15 +500,30 @@ class GraniteASREngineAdapter:
         warnings: List[str] = []
         notes: List[str] = []
         normalized_language = self._normalize_language(req.language)
+        model_name = self.config.get("model_name", "granite-speech-4.1-2b")
+        is_plus = "plus" in str(model_name).lower()
+        
+        diarization_requested = is_plus and req.diarization and req.task == "transcribe"
+        if req.diarization:
+            if not is_plus:
+                warnings.append("Granite diarization skipped: Only 'plus' variants support speaker attribution natively.")
+            elif req.task != "transcribe":
+                warnings.append("Granite diarization skipped: Speaker attribution is only supported for transcription, not translation.")
+
         target_language = self.config.get("asr_translate_target_language", "English")
         translate_instruction_override = str(
             self.config.get("asr_translate_instruction_override", DEFAULT_TRANSLATE_INSTRUCTION_TEMPLATE)
             or DEFAULT_TRANSLATE_INSTRUCTION_TEMPLATE
         ).strip()
         if req.language and str(req.language).lower() != "auto" and not normalized_language:
-            warnings.append(
-                f"Granite received unsupported language hint '{req.language}'. Falling back to Granite auto language handling."
-            )
+            if str(req.language).lower() == "japanese" and "plus" in str(model_name).lower():
+                warnings.append(
+                    "Granite Speech Plus variant does not support Japanese. Falling back to Granite auto language handling."
+                )
+            else:
+                warnings.append(
+                    f"Granite received unsupported language hint '{req.language}'. Falling back to Granite auto language handling."
+                )
         if req.task == "translate":
             self._append_once(
                 notes,
@@ -278,7 +535,8 @@ class GraniteASREngineAdapter:
                     "Granite custom translation instruction override is active on the engine node. This is still wrapped in Granite's chat format, but malformed or weak instructions can fall back to plain transcription.",
                 )
 
-        use_aligner = self._can_use_aligner(req, normalized_language, warnings)
+        use_native_timestamps = is_plus and req.timestamps == "word" and req.task == "transcribe" and not diarization_requested
+        use_aligner = not use_native_timestamps and self._can_use_aligner(req, normalized_language, warnings)
         aligner = self._get_forced_aligner() if use_aligner else None
 
         chunk_size = int(req.chunk_size)
@@ -311,25 +569,58 @@ class GraniteASREngineAdapter:
                     normalized_language,
                     target_language,
                     translate_instruction_override,
+                    timestamps=use_native_timestamps,
+                    diarization=diarization_requested,
                 )
-                if text:
-                    full_text_parts.append(text)
                 if detected_language and not output_language:
                     output_language = detected_language
 
-                if aligner is not None and text:
-                    aligner_language = self._resolve_aligner_language(normalized_language, text, notes)
+                if diarization_requested:
                     chunk_offset = start / sample_rate
-                    segments.extend(
-                        self._align_words(
+                    parsed_text, chunk_segments = parse_granite_diarization(text, chunk_offset, len(chunk_np) / sample_rate)
+                    if parsed_text:
+                        full_text_parts.append(format_granite_diarization_segments(chunk_segments))
+                    
+                    if aligner is not None and parsed_text:
+                        aligner_language = self._resolve_aligner_language(normalized_language, parsed_text, notes)
+                        word_segments = self._align_words(
                             aligner=aligner,
                             chunk_audio=chunk_np,
-                            transcript=text,
+                            transcript=parsed_text,
                             language=aligner_language,
                             chunk_offset=chunk_offset,
                             warnings=warnings,
                         )
-                    )
+                        aligned_words = [seg.words[0] for seg in word_segments if seg.words]
+                        remap_words_to_speaker_segments(chunk_segments, aligned_words)
+                        
+                    for seg in chunk_segments:
+                        if seg.speaker:
+                            seg.text = f"[{seg.speaker}] {seg.text}"
+                    
+                    segments.extend(chunk_segments)
+                elif use_native_timestamps:
+                    chunk_offset = start / sample_rate
+                    parsed_text, chunk_segments = parse_granite_native_timestamps(text, chunk_offset)
+                    if parsed_text:
+                        full_text_parts.append(parsed_text)
+                    segments.extend(chunk_segments)
+                else:
+                    if text:
+                        full_text_parts.append(text)
+                    if aligner is not None and text:
+                        aligner_language = self._resolve_aligner_language(normalized_language, text, notes)
+                        chunk_offset = start / sample_rate
+                        segments.extend(
+                            self._align_words(
+                                aligner=aligner,
+                                chunk_audio=chunk_np,
+                                transcript=text,
+                                language=aligner_language,
+                                chunk_offset=chunk_offset,
+                                warnings=warnings,
+                            )
+                        )
 
                 try:
                     chunk_progress_bar.update(1)
@@ -360,23 +651,54 @@ class GraniteASREngineAdapter:
                 normalized_language,
                 target_language,
                 translate_instruction_override,
+                timestamps=use_native_timestamps,
+                diarization=diarization_requested,
             )
-            if text:
-                full_text_parts.append(text)
             output_language = detected_language or output_language
 
-            if aligner is not None and text:
-                aligner_language = self._resolve_aligner_language(normalized_language, text, notes)
-                segments.extend(
-                    self._align_words(
+            if diarization_requested:
+                parsed_text, chunk_segments = parse_granite_diarization(text, 0.0, len(wav_np) / sample_rate)
+                if parsed_text:
+                    full_text_parts.append(format_granite_diarization_segments(chunk_segments))
+                
+                if aligner is not None and parsed_text:
+                    aligner_language = self._resolve_aligner_language(normalized_language, parsed_text, notes)
+                    word_segments = self._align_words(
                         aligner=aligner,
                         chunk_audio=wav_np,
-                        transcript=text,
+                        transcript=parsed_text,
                         language=aligner_language,
                         chunk_offset=0.0,
                         warnings=warnings,
                     )
-                )
+                    aligned_words = [seg.words[0] for seg in word_segments if seg.words]
+                    remap_words_to_speaker_segments(chunk_segments, aligned_words)
+                    
+                for seg in chunk_segments:
+                    if seg.speaker:
+                        seg.text = f"[{seg.speaker}] {seg.text}"
+                
+                segments.extend(chunk_segments)
+            elif use_native_timestamps:
+                parsed_text, chunk_segments = parse_granite_native_timestamps(text, 0.0)
+                if parsed_text:
+                    full_text_parts.append(parsed_text)
+                segments.extend(chunk_segments)
+            else:
+                if text:
+                    full_text_parts.append(text)
+                if aligner is not None and text:
+                    aligner_language = self._resolve_aligner_language(normalized_language, text, notes)
+                    segments.extend(
+                        self._align_words(
+                            aligner=aligner,
+                            chunk_audio=wav_np,
+                            transcript=text,
+                            language=aligner_language,
+                            chunk_offset=0.0,
+                            warnings=warnings,
+                        )
+                    )
 
         text = " ".join(part for part in full_text_parts if part).strip()
         raw = None

engines/adapters/asr_qwen3_adapter.py

@@ -76,6 +76,8 @@ class Qwen3ASREngineAdapter:
             model_name=model_name,
             model_path=model_id,
             device=device,
+            runtime_mode=self.config.get("runtime_mode", "main_environment"),
+            runtime_profile=self.config.get("runtime_profile"),
             additional_params=additional_params,
         )
 

engines/adapters/audio_cpp_adapter.py

@@ -0,0 +1,372 @@
+"""Adapter between the suite's TTS processors and an audio.cpp session."""
+
+from __future__ import annotations
+
+import json
+import os
+import threading
+from typing import Any, Dict, Mapping, Optional, Tuple
+
+import torch
+
+from utils.audio.audio_hash import generate_stable_audio_component
+from utils.audio.cache import get_audio_cache
+from utils.audio.processing import AudioProcessingUtils
+from utils.voice.reference import effective_voice_audio
+
+
+_CACHE_SAMPLE_RATES: Dict[str, int] = {}
+_CACHE_SAMPLE_RATES_LOCK = threading.Lock()
+
+
+def _get_session(config: Mapping[str, Any]):
+    """Import lazily so the node can still be discovered before optional setup."""
+    from utils.audio_cpp.session import get_audio_cpp_session
+
+    return get_audio_cpp_session(dict(config))
+
+
+def _canonical_json(value: Mapping[str, Any]) -> str:
+    return json.dumps(value, sort_keys=True, separators=(",", ":"), default=str)
+
+
+class AudioCppEngineAdapter:
+    """Build generic ``/v1/tasks/run`` requests and retain their real sample rate."""
+
+    _COMMON_REQUEST_FIELDS = (
+        "temperature",
+        "top_p",
+        "top_k",
+        "repetition_penalty",
+        "max_tokens",
+        "max_steps",
+        "num_inference_steps",
+        "guidance_scale",
+        "speaking_rate",
+    )
+
+    def __init__(self, config: Optional[Dict[str, Any]] = None):
+        self.config = dict(config or {})
+        self.audio_cache = get_audio_cache()
+        self._last_sample_rate: Optional[int] = None
+        self._reference_files: Dict[str, str] = {}
+        self._reference_lock = threading.RLock()
+
+    @property
+    def sample_rate(self) -> Optional[int]:
+        return self._last_sample_rate
+
+    def update_config(self, new_config: Optional[Dict[str, Any]]) -> None:
+        self.config = dict(new_config or {})
+
+    @staticmethod
+    def _reference_text(voice_ref: Any) -> str:
+        if not isinstance(voice_ref, Mapping):
+            return ""
+        return str(
+            voice_ref.get("reference_text")
+            or voice_ref.get("prompt_text")
+            or voice_ref.get("text")
+            or ""
+        ).strip()
+
+    def _materialize_reference(self, voice_ref: Any) -> Tuple[Optional[str], str, str, Optional[str]]:
+        """Return path, transcript, stable hash, and the path that must be removed."""
+        reference_text = AudioCppEngineAdapter._reference_text(voice_ref)
+        if not isinstance(voice_ref, Mapping):
+            return None, reference_text, "default_voice", None
+
+        audio = effective_voice_audio(voice_ref)
+        if audio is None:
+            return None, reference_text, "default_voice", None
+
+        if isinstance(audio, (str, os.PathLike)):
+            path = os.path.abspath(os.path.expanduser(os.fspath(audio)))
+            if not os.path.isfile(path):
+                raise FileNotFoundError(f"audio.cpp reference audio not found: {path}")
+            component = generate_stable_audio_component(audio_file_path=path)
+            return path, reference_text, component, None
+
+        if isinstance(audio, Mapping):
+            waveform = audio.get("waveform")
+            sample_rate = audio.get("sample_rate")
+            audio_dict = dict(audio)
+        elif torch.is_tensor(audio):
+            waveform = audio
+            sample_rate = voice_ref.get("sample_rate")
+            audio_dict = {"waveform": waveform, "sample_rate": sample_rate}
+        else:
+            raise TypeError(f"Unsupported audio.cpp voice reference type: {type(audio).__name__}")
+
+        if not torch.is_tensor(waveform):
+            raise TypeError("audio.cpp reference audio must contain a waveform tensor")
+        if sample_rate is None or int(sample_rate) <= 0:
+            raise ValueError("audio.cpp reference audio must contain a positive sample_rate")
+
+        audio_dict["sample_rate"] = int(sample_rate)
+        component = generate_stable_audio_component(reference_audio=audio_dict)
+        if component not in {"ref_audio_error", "ref_audio_error_not_tensor"}:
+            with self._reference_lock:
+                cached_path = self._reference_files.get(component)
+                if cached_path and os.path.isfile(cached_path):
+                    return cached_path, reference_text, component, None
+                temp_path = os.path.abspath(
+                    AudioProcessingUtils.save_audio_to_temp_file(waveform, int(sample_rate))
+                )
+                self._reference_files[component] = temp_path
+                return temp_path, reference_text, component, None
+
+        # Hash failures must not make unrelated references share one file.
+        temp_path = os.path.abspath(
+            AudioProcessingUtils.save_audio_to_temp_file(waveform, int(sample_rate))
+        )
+        return temp_path, reference_text, component, temp_path
+
+    def close(self) -> None:
+        with self._reference_lock:
+            paths = list(self._reference_files.values())
+            self._reference_files.clear()
+        for path in paths:
+            try:
+                os.remove(path)
+            except FileNotFoundError:
+                pass
+            except OSError:
+                pass
+
+    def __del__(self):
+        try:
+            self.close()
+        except Exception:
+            pass
+
+    def _advanced_options(self) -> Dict[str, Any]:
+        value = self.config.get(
+            "advanced_options",
+            self.config.get("request_options", self.config.get("advanced_json", {})),
+        )
+        if value in (None, ""):
+            return {}
+        if isinstance(value, str):
+            try:
+                value = json.loads(value)
+            except json.JSONDecodeError as exc:
+                raise ValueError(f"Invalid audio.cpp advanced JSON: {exc.msg}") from exc
+        if not isinstance(value, Mapping):
+            raise ValueError("audio.cpp advanced options must be a JSON object")
+        return dict(value)
+
+    def _resolved_task(self, session: Any) -> str:
+        requested = str(self.config.get("task", self.config.get("requested_task", "auto"))).lower()
+        for source in (session, getattr(session, "config", None)):
+            if source is None:
+                continue
+            value = source.get("task") if isinstance(source, Mapping) else getattr(source, "task", None)
+            if str(value).lower() in {"tts", "clon", "vdes"}:
+                return str(value).lower()
+
+        if requested in {"tts", "clon", "vdes"}:
+            return requested
+        if str(self.config.get("connection_mode", "auto")).lower() == "external_server":
+            return "auto"
+        try:
+            from utils.audio_cpp.catalog import resolve_task
+
+            return str(
+                resolve_task(
+                    self.config.get("family", ""),
+                    self.config.get("package_id", ""),
+                    requested="auto",
+                )
+            ).lower()
+        except (ImportError, KeyError, TypeError, ValueError):
+            return "tts"
+
+    def _build_request(
+        self,
+        text: str,
+        voice_path: Optional[str],
+        reference_text: str,
+        seed: int,
+        advanced: Dict[str, Any],
+        task: str,
+    ) -> Dict[str, Any]:
+        request: Dict[str, Any] = {"text": text, "seed": str(int(seed)), "options": advanced}
+        del task  # The persistent session owns its one configured model/task.
+
+        language = str(self.config.get("language", "")).strip()
+        if language and language.lower() not in {"auto", "none"}:
+            request["language"] = language
+        voice_id = str(self.config.get("voice_id", self.config.get("voice", ""))).strip()
+        if voice_id:
+            request["voice_id"] = voice_id
+        if voice_path:
+            request["voice_ref"] = voice_path
+        if reference_text:
+            request["reference_text"] = reference_text
+        instruct = str(self.config.get("instruct", "")).strip()
+        if instruct:
+            request["instruct"] = instruct
+
+        for key in self._COMMON_REQUEST_FIELDS:
+            value = self.config.get(key)
+            if value is not None and value != "":
+                request[key] = value
+        return request
+
+    def _cache_key(
+        self,
+        text: str,
+        audio_component: str,
+        reference_text: str,
+        seed: int,
+        task: str,
+        advanced: Dict[str, Any],
+        character_name: Optional[str],
+        session: Any,
+    ) -> str:
+        session_config = getattr(session, "config", {})
+        if not isinstance(session_config, Mapping):
+            session_config = {}
+        session_family = getattr(session, "family", None) or session_config.get(
+            "family", self.config.get("family", "")
+        )
+        session_model_id = getattr(session, "model_id", None) or session_config.get(
+            "model_id", self.config.get("model_id", "")
+        )
+        # Owned servers use a random loopback port on every restart; that port is
+        # transport state, not model identity. External endpoints are stable and
+        # must participate in the cache key.
+        if bool(getattr(session, "owned", False)):
+            session_endpoint = ""
+        else:
+            session_endpoint = getattr(session, "endpoint", None) or self.config.get(
+                "server_url", self.config.get("external_server_url", "")
+            )
+        extra_identity = {
+            "options": advanced,
+            "speaking_rate": self.config.get("speaking_rate"),
+            "connection_mode": self.config.get("connection_mode", "auto"),
+            "server_url": session_endpoint,
+            "binary_path": session_config.get("binary_path", self.config.get("binary_path", "")),
+            "backend": session_config.get("backend", self.config.get("backend", "")),
+            "device": session_config.get("device", self.config.get("device", "")),
+            "load_options": session_config.get("load_options", self.config.get("load_options", {})),
+            "session_options": session_config.get(
+                "session_options", self.config.get("session_options", {})
+            ),
+            "default_request_options": session_config.get(
+                "default_request_options", self.config.get("default_request_options", {})
+            ),
+        }
+        return self.audio_cache.generate_cache_key(
+            "audio_cpp",
+            text=text,
+            audio_component=audio_component,
+            reference_text=reference_text,
+            family=session_family,
+            package_id=session_config.get("package_id", self.config.get("package_id", "")),
+            model_path=session_config.get("model_path", self.config.get("model_path", "")),
+            model_id=session_model_id,
+            task=task,
+            language=self.config.get("language", ""),
+            voice_id=self.config.get("voice_id", self.config.get("voice", "")),
+            instruct=self.config.get("instruct", ""),
+            temperature=self.config.get("temperature"),
+            top_p=self.config.get("top_p"),
+            top_k=self.config.get("top_k"),
+            repetition_penalty=self.config.get("repetition_penalty"),
+            max_tokens=self.config.get("max_tokens"),
+            max_steps=self.config.get("max_steps"),
+            num_inference_steps=self.config.get("num_inference_steps"),
+            guidance_scale=self.config.get("guidance_scale"),
+            seed=int(seed),
+            request_options=_canonical_json(extra_identity),
+            character=character_name or "narrator",
+        )
+
+    @staticmethod
+    def _normalize_result(result: Any) -> Tuple[torch.Tensor, int]:
+        waveform = result.get("waveform") if isinstance(result, Mapping) else getattr(result, "waveform", None)
+        sample_rate = result.get("sample_rate") if isinstance(result, Mapping) else getattr(result, "sample_rate", None)
+
+        if waveform is None:
+            named = result.get("named_audio", {}) if isinstance(result, Mapping) else getattr(result, "named_audio", {})
+            values = list(named.values()) if isinstance(named, Mapping) else list(named or [])
+            if len(values) == 1:
+                item = values[0]
+                waveform = item.get("waveform") if isinstance(item, Mapping) else getattr(item, "waveform", None)
+                sample_rate = sample_rate or (item.get("sample_rate") if isinstance(item, Mapping) else getattr(item, "sample_rate", None))
+
+        if waveform is None:
+            raise RuntimeError("audio.cpp returned no primary audio output")
+        if not torch.is_tensor(waveform):
+            waveform = torch.as_tensor(waveform, dtype=torch.float32)
+        waveform = waveform.detach().to(device="cpu", dtype=torch.float32)
+        if waveform.dim() == 1:
+            waveform = waveform.unsqueeze(0)
+        elif waveform.dim() == 3 and waveform.shape[0] == 1:
+            waveform = waveform.squeeze(0)
+        if waveform.dim() != 2:
+            raise ValueError(f"audio.cpp waveform must be [channels, samples], got {tuple(waveform.shape)}")
+        if sample_rate is None or int(sample_rate) <= 0:
+            raise ValueError("audio.cpp returned an invalid sample rate")
+        return waveform.contiguous(), int(sample_rate)
+
+    def generate_single(
+        self,
+        text: str,
+        voice_ref: Optional[Dict[str, Any]] = None,
+        seed: int = 0,
+        enable_audio_cache: bool = True,
+        character_name: Optional[str] = None,
+    ) -> Tuple[torch.Tensor, int]:
+        stripped = str(text or "").strip()
+        if not stripped:
+            if self._last_sample_rate is None:
+                raise ValueError("audio.cpp cannot determine a sample rate for empty text")
+            return torch.zeros(1, 0, dtype=torch.float32), self._last_sample_rate
+
+        session = _get_session(self.config)
+        task = self._resolved_task(session)
+        advanced = self._advanced_options()
+        cleanup_path: Optional[str] = None
+        try:
+            voice_path, reference_text, audio_component, cleanup_path = self._materialize_reference(voice_ref)
+            cache_key = self._cache_key(
+                stripped,
+                audio_component,
+                reference_text,
+                seed,
+                task,
+                advanced,
+                character_name,
+                session,
+            )
+            if enable_audio_cache:
+                cached = self.audio_cache.get_cached_audio(cache_key)
+                with _CACHE_SAMPLE_RATES_LOCK:
+                    cached_rate = _CACHE_SAMPLE_RATES.get(cache_key)
+                if cached is not None and cached_rate is not None:
+                    self._last_sample_rate = cached_rate
+                    return cached[0].clone(), cached_rate
+
+            request = self._build_request(stripped, voice_path, reference_text, seed, advanced, task)
+            waveform, sample_rate = self._normalize_result(session.run(request))
+            self._last_sample_rate = sample_rate
+            if enable_audio_cache:
+                duration = waveform.shape[-1] / sample_rate
+                self.audio_cache.cache_audio(cache_key, waveform, duration)
+                with _CACHE_SAMPLE_RATES_LOCK:
+                    _CACHE_SAMPLE_RATES[cache_key] = sample_rate
+            return waveform, sample_rate
+        finally:
+            if cleanup_path:
+                try:
+                    os.remove(cleanup_path)
+                except FileNotFoundError:
+                    pass
+
+
+# Short alias for callers that do not use the older ``EngineAdapter`` suffix.
+AudioCppAdapter = AudioCppEngineAdapter

engines/adapters/audio_cpp_vc_adapter.py

@@ -0,0 +1,111 @@
+"""audio.cpp adapter for the Suite's unified Voice Changer node."""
+
+from __future__ import annotations
+
+import json
+import os
+from typing import Any, Dict, Mapping
+
+import torch
+
+from engines.adapters.audio_cpp_adapter import AudioCppEngineAdapter
+from utils.audio.processing import AudioProcessingUtils
+
+
+def _advanced_options(config: Mapping[str, Any]) -> Dict[str, Any]:
+    value = config.get("advanced_options", config.get("request_options", {}))
+    if value in (None, ""):
+        return {}
+    if isinstance(value, str):
+        try:
+            value = json.loads(value)
+        except json.JSONDecodeError as exc:
+            raise ValueError(f"Invalid audio.cpp advanced JSON: {exc.msg}") from exc
+    if not isinstance(value, Mapping):
+        raise ValueError("audio.cpp advanced options must be a JSON object")
+    return dict(value)
+
+
+def _materialize(audio: Mapping[str, Any], label: str) -> str:
+    waveform = audio.get("waveform")
+    sample_rate = audio.get("sample_rate")
+    if not torch.is_tensor(waveform):
+        raise TypeError(f"audio.cpp {label} must contain a waveform tensor")
+    if sample_rate is None or int(sample_rate) <= 0:
+        raise ValueError(f"audio.cpp {label} must contain a positive sample_rate")
+    return os.path.abspath(
+        AudioProcessingUtils.save_audio_to_temp_file(waveform, int(sample_rate))
+    )
+
+
+class AudioCppVoiceConversionAdapter:
+    """Convert source audio toward a target reference using an audio.cpp VC task."""
+
+    def __init__(self, config: Dict[str, Any]):
+        self.config = dict(config)
+
+    def _session_config(self) -> Dict[str, Any]:
+        config = dict(self.config)
+        if str(config.get("connection_mode", "auto")).lower() != "external_server":
+            config["requested_task"] = "vc"
+            config["task"] = "vc"
+        return config
+
+    def convert_voice(
+        self,
+        source_audio: Dict[str, Any],
+        target_audio: Dict[str, Any],
+        refinement_passes: int = 1,
+    ) -> tuple[Dict[str, Any], str]:
+        from utils.audio_cpp.session import get_audio_cpp_session
+
+        config = self._session_config()
+        family = str(config.get("family", "")).strip()
+        passes = max(1, int(refinement_passes))
+        current = source_audio
+        output_rate = int(source_audio["sample_rate"])
+
+        session = get_audio_cpp_session(config)
+        if str(getattr(session, "task", "vc")) != "vc":
+            raise ValueError(
+                f"audio.cpp model '{session.model_id}' is configured for task "
+                f"'{session.task}', not voice conversion"
+            )
+
+        for pass_index in range(passes):
+            source_path = _materialize(current, "source audio")
+            target_path = _materialize(target_audio, "target reference audio")
+            try:
+                request = {
+                    "audio": source_path,
+                    "voice_ref": target_path,
+                    "source_audio": source_path,
+                    "target_voice": target_path,
+                    "options": _advanced_options(config),
+                }
+                print(
+                    f"🔄 audio.cpp VC: {family or 'external model'} pass "
+                    f"{pass_index + 1}/{passes}..."
+                )
+                result = session.run(request)
+                waveform, output_rate = AudioCppEngineAdapter._normalize_result(result)
+                current = {"waveform": waveform.unsqueeze(0), "sample_rate": output_rate}
+            finally:
+                for path in (source_path, target_path):
+                    try:
+                        os.remove(path)
+                    except FileNotFoundError:
+                        pass
+
+        info = (
+            f"Model family: {family or getattr(session, 'family', 'external')}\n"
+            f"Model ID: {session.model_id}\n"
+            f"Task: voice conversion\n"
+            f"Refinement passes: {passes}\n"
+            f"Output sample rate: {output_rate} Hz\n"
+            "Conversion completed successfully"
+        )
+        return current, info
+
+
+__all__ = ["AudioCppVoiceConversionAdapter"]

engines/adapters/cosyvoice_adapter.py

@@ -23,6 +23,7 @@ from engines.cosyvoice.cosyvoice import CosyVoiceEngine
 from engines.cosyvoice.cosyvoice_downloader import cosyvoice_downloader
 from utils.text.character_parser import character_parser
 from utils.voice.discovery import get_character_mapping, get_available_characters
+from utils.voice.character_logging import resolved_character_label
 from utils.audio.cache import get_audio_cache
 
 
@@ -336,7 +337,7 @@ class CosyVoiceAdapter:
                     speaker_audio = char_audio
                     if char_text:
                         reference_text = char_text
-                    print(f"📖 Using character voice '{character_name}'")
+                    print(f"📖 Using character voice '{resolved_character_label(character_name, speaker_audio)}'")
 
             # Generate cache key for this segment
             segment_cache_key = self._generate_cache_key(
@@ -352,7 +353,7 @@ class CosyVoiceAdapter:
             # Check cache first
             cached_segment_audio = self.audio_cache.get_cached_audio(segment_cache_key)
             if cached_segment_audio:
-                print(f"💾 Using cached CosyVoice3 segment for '{character_name}'")
+                print(f"💾 Using cached CosyVoice3 segment for '{resolved_character_label(character_name, speaker_audio)}'")
                 segment_audio = cached_segment_audio[0]
             else:
                 # Convert CosyVoice paralinguistic tags from <tag> to [tag]

engines/adapters/dots_tts_adapter.py

@@ -0,0 +1,221 @@
+"""
+Dots TTS engine adapter.
+
+Translates TTS Audio Suite processor calls into the official dots.tts runtime.
+"""
+
+import os
+import sys
+from typing import Any, Dict, Optional, Tuple
+
+import torch
+
+current_dir = os.path.dirname(__file__)
+engines_dir = os.path.dirname(current_dir)
+project_root = os.path.dirname(engines_dir)
+if project_root not in sys.path:
+    sys.path.insert(0, project_root)
+
+from utils.audio.audio_hash import generate_stable_audio_component
+from utils.audio.cache import get_audio_cache
+from utils.audio.processing import AudioProcessingUtils
+from utils.models.factory_config import ModelLoadConfig
+from utils.models.language_mapper import resolve_language_alias
+from utils.voice.reference import effective_voice_audio
+from engines.dots_tts.languages import normalize_dots_language
+
+
+class DotsTTSEngineAdapter:
+    """Adapter for official Dots TTS inference."""
+
+    SAMPLE_RATE = 48000
+
+    def __init__(self, config: Optional[Dict[str, Any]] = None):
+        self.config = config.copy() if config else {}
+        self.audio_cache = get_audio_cache()
+        self._last_config: Optional[ModelLoadConfig] = None
+        self._load_signature = None
+
+    def update_config(self, new_config: Dict[str, Any]):
+        self.config = new_config.copy() if new_config else {}
+
+    def _build_load_signature(self) -> Tuple[Any, ...]:
+        return (
+            self.config.get("model_variant", "dots.tts-soar"),
+            self.config.get("device", "auto"),
+            self.config.get("precision", "auto"),
+            bool(self.config.get("optimize", False)),
+            int(self.config.get("max_generate_length", 500)),
+        )
+
+    def load_model(
+        self,
+        model_variant: str,
+        device: str = "auto",
+        precision: str = "auto",
+        optimize: bool = False,
+        max_generate_length: int = 500,
+    ):
+        from utils.models.unified_model_interface import unified_model_interface
+
+        config = ModelLoadConfig(
+            engine_name="dots_tts",
+            model_type="tts",
+            model_name=model_variant,
+            model_path=model_variant,
+            device=device,
+            additional_params={
+                "precision": precision,
+                "optimize": bool(optimize),
+                "max_generate_length": int(max_generate_length),
+            },
+        )
+        self._last_config = config
+        unified_model_interface.load_model(config)
+
+    def _ensure_model_loaded(self):
+        signature = self._build_load_signature()
+        if signature == self._load_signature and self._last_config is not None:
+            return
+
+        self.load_model(
+            model_variant=self.config.get("model_variant", "dots.tts-soar"),
+            device=self.config.get("device", "auto"),
+            precision=self.config.get("precision", "auto"),
+            optimize=self.config.get("optimize", False),
+            max_generate_length=self.config.get("max_generate_length", 500),
+        )
+        self._load_signature = signature
+
+    def _get_engine(self):
+        if self._last_config is None:
+            self._ensure_model_loaded()
+        from utils.models.unified_model_interface import unified_model_interface
+
+        return unified_model_interface.load_model(self._last_config)
+
+    def _extract_voice_reference(self, voice_ref: Optional[Dict[str, Any]]) -> Tuple[Optional[str], str, str]:
+        if not voice_ref or not isinstance(voice_ref, dict):
+            return None, "", "default_voice"
+
+        prompt_text = (
+            voice_ref.get("reference_text")
+            or voice_ref.get("prompt_text")
+            or voice_ref.get("text")
+            or ""
+        ).strip()
+
+        ref_audio = effective_voice_audio(voice_ref)
+
+        if ref_audio is None:
+            return None, prompt_text, "default_voice"
+
+        if isinstance(ref_audio, str):
+            return ref_audio, prompt_text, generate_stable_audio_component(audio_file_path=ref_audio)
+
+        if isinstance(ref_audio, dict) and "waveform" in ref_audio:
+            temp_path = AudioProcessingUtils.save_audio_to_temp_file(
+                ref_audio["waveform"],
+                ref_audio.get("sample_rate", self.SAMPLE_RATE),
+            )
+            return temp_path, prompt_text, generate_stable_audio_component(reference_audio=ref_audio)
+
+        if torch.is_tensor(ref_audio):
+            sample_rate = int(voice_ref.get("sample_rate", self.SAMPLE_RATE))
+            audio_dict = {"waveform": ref_audio, "sample_rate": sample_rate}
+            temp_path = AudioProcessingUtils.save_audio_to_temp_file(ref_audio, sample_rate)
+            return temp_path, prompt_text, generate_stable_audio_component(reference_audio=audio_dict)
+
+        raise TypeError(f"Unsupported Dots TTS voice reference type: {type(ref_audio)}")
+
+    @staticmethod
+    def _normalize_language(language: Optional[str]) -> Optional[str]:
+        if language is None:
+            return None
+        normalized = str(language).strip()
+        if not normalized:
+            return None
+        normalized_language = normalize_dots_language(normalized)
+        if normalized_language is not None:
+            return normalized_language
+        resolved = resolve_language_alias(normalized)
+        return normalize_dots_language(resolved)
+
+    def generate_single(
+        self,
+        text: str,
+        voice_ref: Optional[Dict[str, Any]],
+        seed: int = 0,
+        enable_audio_cache: bool = True,
+        character_name: Optional[str] = None,
+    ) -> torch.Tensor:
+        stripped = (text or "").strip()
+        if not stripped:
+            return torch.zeros(1, 0, dtype=torch.float32)
+
+        self._ensure_model_loaded()
+        engine = self._get_engine()
+
+        prompt_audio_path, prompt_text, audio_component = self._extract_voice_reference(voice_ref)
+        language = self._normalize_language(self.config.get("language", "auto"))
+        model_variant = self.config.get("model_variant", "dots.tts-soar")
+        template_name = self.config.get("template_name", "tts")
+        num_steps = int(self.config.get("num_steps", 10))
+        guidance_scale = float(self.config.get("guidance_scale", 1.2))
+        speaker_scale = float(self.config.get("speaker_scale", 1.5))
+        normalize_text = bool(self.config.get("normalize_text", False))
+        precision = self.config.get("precision", "auto")
+        optimize = bool(self.config.get("optimize", False))
+        max_generate_length = int(self.config.get("max_generate_length", 500))
+
+        cache_key = None
+        if enable_audio_cache:
+            cache_key = self.audio_cache.generate_cache_key(
+                "dots_tts",
+                text=stripped,
+                audio_component=audio_component,
+                prompt_text=prompt_text,
+                model_variant=model_variant,
+                language=language or "none",
+                template_name=template_name,
+                num_steps=num_steps,
+                guidance_scale=guidance_scale,
+                speaker_scale=speaker_scale,
+                normalize_text=normalize_text,
+                max_generate_length=max_generate_length,
+                precision=precision,
+                optimize=optimize,
+                seed=int(seed or 0),
+                device=self.config.get("device", "auto"),
+                character=character_name or "narrator",
+            )
+            cached_audio = self.audio_cache.get_cached_audio(cache_key)
+            if cached_audio:
+                print(f"💾 Using cached Dots TTS audio for '{character_name or 'narrator'}': '{stripped[:30]}...'")
+                return cached_audio[0]
+
+        result = engine.generate(
+            text=stripped,
+            prompt_audio_path=prompt_audio_path,
+            prompt_text=prompt_text or None,
+            language=language,
+            num_steps=num_steps,
+            guidance_scale=guidance_scale,
+            speaker_scale=speaker_scale,
+            normalize_text=normalize_text,
+            template_name=template_name,
+            seed=int(seed or 0),
+        )
+
+        audio_tensor = result["audio"]
+        if not isinstance(audio_tensor, torch.Tensor):
+            audio_tensor = torch.tensor(audio_tensor, dtype=torch.float32)
+        audio_tensor = audio_tensor.detach().float().cpu()
+        if audio_tensor.dim() == 1:
+            audio_tensor = audio_tensor.unsqueeze(0)
+
+        if enable_audio_cache and cache_key:
+            duration = self.audio_cache._calculate_duration(audio_tensor, "dots_tts")
+            self.audio_cache.cache_audio(cache_key, audio_tensor, duration)
+
+        return audio_tensor

engines/adapters/dramabox_adapter.py

@@ -0,0 +1,300 @@
+"""Adapter between unified TTS processing and official DramaBox inference."""
+
+import math
+import os
+from typing import Any, Dict, Optional, Tuple
+
+import torch
+import torchaudio
+
+from utils.audio.audio_hash import generate_stable_audio_component
+from utils.audio.cache import get_audio_cache
+from utils.audio.processing import AudioProcessingUtils
+from utils.models.factory_config import ModelLoadConfig
+from utils.voice.reference import effective_voice_audio
+
+
+class DramaBoxEngineAdapter:
+    """Translate suite voice/config/cache data into DramaBox calls."""
+
+    SAMPLE_RATE = 48000
+    SILENCE_RMS_THRESHOLD = 1e-3
+    SILENCE_PEAK_THRESHOLD = 2e-2
+
+    def __init__(self, config: Optional[Dict[str, Any]] = None):
+        self.config = config.copy() if config else {}
+        self.audio_cache = get_audio_cache()
+        self._last_config: Optional[ModelLoadConfig] = None
+        self._load_signature = None
+        self._lora_signature = None
+        self.last_generation_status: Dict[str, Any] = {"near_silent": False}
+
+    def update_config(self, new_config: Dict[str, Any]):
+        self.config = new_config.copy() if new_config else {}
+
+    @staticmethod
+    def _lora_revision(path: Any) -> str:
+        """Return a cheap cache token that changes when a managed adapter is replaced."""
+        value = str(path or "").strip()
+        if not value:
+            return ""
+        try:
+            candidate = os.path.abspath(os.path.expanduser(value))
+            if os.path.isfile(candidate):
+                stat = os.stat(candidate)
+                return f"{candidate}:{stat.st_size}:{stat.st_mtime_ns}"
+            if os.path.isdir(candidate):
+                entries = []
+                for item in os.listdir(candidate):
+                    if not item.endswith(".safetensors"):
+                        continue
+                    item_path = os.path.join(candidate, item)
+                    stat = os.stat(item_path)
+                    entries.append(f"{item}:{stat.st_size}:{stat.st_mtime_ns}")
+                return f"{candidate}|{'|'.join(sorted(entries))}"
+        except OSError:
+            pass
+        return value
+
+    @classmethod
+    def _warn_if_near_silent(
+        cls,
+        audio: torch.Tensor,
+        *,
+        character_name: Optional[str],
+        seed: int,
+        cached: bool = False,
+    ) -> Optional[Dict[str, Any]]:
+        """Warn about clearly near-silent model output without altering it."""
+        if not isinstance(audio, torch.Tensor) or audio.numel() == 0:
+            return None
+
+        samples = torch.nan_to_num(audio.detach().float().cpu())
+        rms = float(samples.square().mean().sqrt())
+        peak = float(samples.abs().max())
+        if rms >= cls.SILENCE_RMS_THRESHOLD or peak >= cls.SILENCE_PEAK_THRESHOLD:
+            return None
+
+        rms_db = 20.0 * math.log10(max(rms, 1e-12))
+        peak_db = 20.0 * math.log10(max(peak, 1e-12))
+        source = "cached " if cached else ""
+        print(
+            f"\n⚠️ DramaBox generated a near-silent {source}segment for "
+            f"'{character_name or 'narrator'}' "
+            f"(RMS {rms_db:.1f} dBFS, peak {peak_db:.1f} dBFS)."
+        )
+        print(
+            "⚠️ This can depend on generation duration, reference duration, "
+            "reference audio, guidance settings, and seed."
+        )
+        print(
+            "⚠️ Try changing those parameters for this segment; another seed "
+            "may help, but is not guaranteed to fix it.\n"
+        )
+        return {
+            "near_silent": True,
+            "character": character_name or "narrator",
+            "seed": int(seed),
+            "rms_dbfs": rms_db,
+            "peak_dbfs": peak_db,
+            "cached": bool(cached),
+        }
+
+    def _build_load_signature(self) -> Tuple[Any, ...]:
+        """Identity of the expensive base runtime, excluding live LoRA state."""
+        return (
+            self.config.get("model_name", "DramaBox"),
+            self.config.get("device", "auto"),
+            self.config.get("precision", "auto"),
+            self.config.get("memory_mode", "fast"),
+            self.config.get("transformer_quantization", "none"),
+            bool(self.config.get("compile_model", False)),
+        )
+
+    def _build_lora_signature(self) -> Tuple[Any, ...]:
+        path = self.config.get("lora_path", "")
+        return (
+            str(path or "").strip(),
+            self._lora_revision(path),
+            float(self.config.get("lora_strength", 1.0)),
+        )
+
+    def _ensure_model_loaded(self):
+        signature = self._build_load_signature()
+        from utils.models.unified_model_interface import unified_model_interface
+
+        if signature != self._load_signature or self._last_config is None:
+            self._last_config = ModelLoadConfig(
+                engine_name="dramabox",
+                model_type="tts",
+                model_name=self.config.get("model_name", "DramaBox"),
+                device=self.config.get("device", "auto"),
+                additional_params={
+                    "precision": self.config.get("precision", "auto"),
+                    "memory_mode": self.config.get("memory_mode", "fast"),
+                    "transformer_quantization": self.config.get(
+                        "transformer_quantization", "none"
+                    ),
+                    "compile_model": bool(self.config.get("compile_model", False)),
+                },
+            )
+            self._load_signature = signature
+            self._lora_signature = None
+
+        engine = unified_model_interface.load_model(self._last_config)
+        lora_signature = self._build_lora_signature()
+        if lora_signature != self._lora_signature:
+            lora_path, lora_revision, lora_strength = lora_signature
+            engine.set_lora(
+                lora_path=lora_path,
+                strength=lora_strength,
+                revision=lora_revision,
+            )
+            self._lora_signature = lora_signature
+        return engine
+
+    def _get_engine(self):
+        return self._ensure_model_loaded()
+
+    def _extract_voice_reference(
+        self, voice_ref: Optional[Dict[str, Any]]
+    ) -> Tuple[Optional[str], str, bool]:
+        if not isinstance(voice_ref, dict):
+            return None, "default_voice", False
+
+        audio = effective_voice_audio(voice_ref)
+        if audio is None:
+            return None, "default_voice", False
+        if isinstance(audio, str):
+            return (
+                audio,
+                generate_stable_audio_component(audio_file_path=audio),
+                False,
+            )
+        if isinstance(audio, dict) and "waveform" in audio:
+            path = AudioProcessingUtils.save_audio_to_temp_file(
+                audio["waveform"], audio.get("sample_rate", self.SAMPLE_RATE)
+            )
+            return path, generate_stable_audio_component(reference_audio=audio), True
+        if torch.is_tensor(audio):
+            sample_rate = int(voice_ref.get("sample_rate", self.SAMPLE_RATE))
+            audio_dict = {"waveform": audio, "sample_rate": sample_rate}
+            path = AudioProcessingUtils.save_audio_to_temp_file(audio, sample_rate)
+            return (
+                path,
+                generate_stable_audio_component(reference_audio=audio_dict),
+                True,
+            )
+        raise TypeError(f"Unsupported DramaBox voice reference: {type(audio)}")
+
+    def generate_single(
+        self,
+        text: str,
+        voice_ref: Optional[Dict[str, Any]],
+        seed: int = 42,
+        enable_audio_cache: bool = True,
+        character_name: Optional[str] = None,
+    ) -> torch.Tensor:
+        prompt = (text or "").strip()
+        if not prompt:
+            return torch.zeros(1, 0, dtype=torch.float32)
+
+        voice_path, audio_component, remove_voice_path = self._extract_voice_reference(
+            voice_ref
+        )
+        cfg_scale = float(self.config.get("cfg_scale", 2.5))
+        stg_scale = float(self.config.get("stg_scale", 1.5))
+        duration_multiplier = float(self.config.get("duration_multiplier", 1.1))
+        gen_duration = float(self.config.get("gen_duration", 0.0))
+        ref_duration = float(self.config.get("ref_duration", 10.0))
+        rescale_scale = self.config.get("rescale_scale", "auto")
+        watermark = bool(self.config.get("watermark", False))
+        negative_prompt = str(self.config.get("negative_prompt", ""))
+        model_name = self.config.get("model_name", "DramaBox")
+
+        cache_key = None
+        if enable_audio_cache:
+            cache_key = self.audio_cache.generate_cache_key(
+                "dramabox",
+                text=prompt,
+                audio_component=audio_component,
+                model_name=model_name,
+                cfg_scale=cfg_scale,
+                stg_scale=stg_scale,
+                duration_multiplier=duration_multiplier,
+                gen_duration=gen_duration,
+                ref_duration=ref_duration,
+                rescale_scale=rescale_scale,
+                watermark=watermark,
+                prompt_template=str(
+                    self.config.get("prompt_template", '"{seg}"')
+                ),
+                negative_prompt=negative_prompt,
+                precision=self.config.get("precision", "auto"),
+                transformer_quantization=self.config.get(
+                    "transformer_quantization", "none"
+                ),
+                memory_mode=self.config.get("memory_mode", "fast"),
+                compile_model=bool(self.config.get("compile_model", False)),
+                lora_path=self.config.get("lora_path", ""),
+                lora_strength=float(self.config.get("lora_strength", 1.0)),
+                lora_revision=self._lora_revision(self.config.get("lora_path", "")),
+                seed=int(seed),
+                character=character_name or "narrator",
+            )
+            cached = self.audio_cache.get_cached_audio(cache_key)
+            if cached:
+                print(
+                    f"💾 Using cached DramaBox audio for "
+                    f"'{character_name or 'narrator'}': '{prompt[:30]}...'"
+                )
+                self.last_generation_status = self._warn_if_near_silent(
+                    cached[0],
+                    character_name=character_name,
+                    seed=int(seed),
+                    cached=True,
+                ) or {"near_silent": False}
+                return cached[0]
+
+        try:
+            result = self._get_engine().generate(
+                prompt=prompt,
+                voice_ref_path=voice_path,
+                cfg_scale=cfg_scale,
+                stg_scale=stg_scale,
+                duration_multiplier=duration_multiplier,
+                gen_duration=gen_duration,
+                ref_duration=ref_duration,
+                rescale_scale=rescale_scale,
+                watermark=watermark,
+                negative_prompt=negative_prompt,
+                seed=int(seed),
+            )
+        finally:
+            if remove_voice_path and voice_path:
+                try:
+                    os.unlink(voice_path)
+                except OSError:
+                    pass
+        audio = result["audio"]
+        if not isinstance(audio, torch.Tensor):
+            audio = torch.tensor(audio, dtype=torch.float32)
+        audio = audio.detach().float().cpu()
+        if audio.dim() == 1:
+            audio = audio.unsqueeze(0)
+
+        if int(result.get("sample_rate", self.SAMPLE_RATE)) != self.SAMPLE_RATE:
+            audio = torchaudio.functional.resample(
+                audio, int(result["sample_rate"]), self.SAMPLE_RATE
+            )
+
+        self.last_generation_status = self._warn_if_near_silent(
+            audio,
+            character_name=character_name,
+            seed=int(seed),
+        ) or {"near_silent": False}
+
+        if enable_audio_cache and cache_key:
+            duration = audio.shape[-1] / self.SAMPLE_RATE
+            self.audio_cache.cache_audio(cache_key, audio, duration)
+        return audio

engines/adapters/fish_audio_s2_adapter.py

@@ -0,0 +1,133 @@
+"""Adapter between suite processors and the isolated official Fish S2 runtime."""
+
+from typing import Any, Dict, Optional
+
+import torch
+
+from engines.fish_audio_s2.downloader import FishAudioS2Downloader
+from utils.audio.audio_hash import generate_stable_audio_component
+from utils.audio.cache import get_audio_cache
+from utils.audio.processing import AudioProcessingUtils
+from utils.models.factory_config import ModelLoadConfig
+from utils.text.fish_audio_s2_tags import translate_fish_s2_inline_tags
+from utils.voice.character_logging import resolved_voice_name
+from utils.voice.reference import effective_voice_audio
+
+
+class FishAudioS2Adapter:
+    SAMPLE_RATE = 44100
+
+    def __init__(self, config: Optional[Dict[str, Any]] = None):
+        self.config = dict(config or {})
+        self.audio_cache = get_audio_cache()
+        self._model_config = None
+
+    def update_config(self, config):
+        self.config = dict(config or {})
+
+    def _engine(self):
+        from utils.models.unified_model_interface import unified_model_interface
+        model_selection = self.config.get("model_variant", "s2-pro")
+        quantization = self.config.get("quantization", "none")
+        model_path = FishAudioS2Downloader.resolve_model_path(model_selection)
+        model_variant = FishAudioS2Downloader.resolve_model_variant(model_selection, model_path)
+        self._model_config = ModelLoadConfig(
+            engine_name="fish_audio_s2", model_type="tts", model_name=model_variant,
+            model_path=model_path, device=self.config.get("device", "auto"),
+            runtime_mode="isolated",
+            additional_params={
+                "model_variant": model_variant,
+                "quantization": quantization,
+                "precision": self.config.get("precision", "bfloat16"),
+                "compile": bool(self.config.get("compile", False)),
+                "context_length": int(self.config.get("context_length", 8192)),
+            },
+        )
+        return unified_model_interface.load_model(self._model_config)
+
+    def _reference(self, voice_ref):
+        if not isinstance(voice_ref, dict):
+            return None, "", "default_voice"
+        text = (voice_ref.get("reference_text") or voice_ref.get("prompt_text") or "").strip()
+        effective_audio = effective_voice_audio(voice_ref)
+        path = effective_audio if isinstance(effective_audio, str) else None
+        if isinstance(effective_audio, dict):
+            path = AudioProcessingUtils.save_audio_to_temp_file(
+                effective_audio["waveform"], effective_audio.get("sample_rate", self.SAMPLE_RATE)
+            )
+        elif torch.is_tensor(effective_audio):
+            path = AudioProcessingUtils.save_audio_to_temp_file(
+                effective_audio, voice_ref.get("sample_rate", self.SAMPLE_RATE)
+            )
+        component = generate_stable_audio_component(audio_file_path=path) if path else "default_voice"
+        return path, text, component
+
+    def generate_single(self, text, voice_ref, seed=0, enable_audio_cache=True, character_name=None):
+        return self.generate_dialogue(
+            [(0, text)], [voice_ref], seed, enable_audio_cache,
+            cache_character=character_name or "narrator",
+        )
+
+    def generate_dialogue(self, turns, voice_refs, seed=0, enable_audio_cache=True,
+                          cache_character="native_dialogue"):
+        formatted_turns = []
+        for speaker_index, turn_text in turns:
+            clean_text = translate_fish_s2_inline_tags((turn_text or "").strip())
+            if clean_text:
+                formatted_turns.append(f"<|speaker:{speaker_index}|>{clean_text}")
+        text = "\n".join(formatted_turns)
+        if not text:
+            return torch.zeros(1, 0)
+
+        references = []
+        reference_labels = []
+        components = []
+        reference_texts = []
+        for speaker_index, voice_ref in enumerate(voice_refs):
+            ref_path, ref_text, component = self._reference(voice_ref)
+            if ref_path:
+                if not ref_text:
+                    raise ValueError("Fish S2 native speakers require exact reference transcripts")
+                references.append({
+                    "audio_path": ref_path,
+                    "text": f"<|speaker:{speaker_index}|>{ref_text}",
+                })
+                reference_labels.append(
+                    f"local Speaker {speaker_index + 1}={resolved_voice_name(voice_ref)}"
+                )
+            components.append(component)
+            reference_texts.append(ref_text)
+        if references:
+            print(f"🎤 Fish local reference order: {', '.join(reference_labels)}")
+
+        params = {
+            "model_variant": self.config.get("model_variant", "s2-pro"),
+            "quantization": self.config.get("quantization", "none"),
+            "multi_speaker_mode": self.config.get("multi_speaker_mode", "Native Multi-Speaker"),
+            "seed": int(seed), "normalize": bool(self.config.get("normalize", True)),
+            "chunk_length": int(self.config.get("native_chunk_length", 200)),
+            "max_new_tokens": int(self.config.get("max_new_tokens", 1024)),
+            "top_p": float(self.config.get("top_p", 0.8)),
+            "repetition_penalty": float(self.config.get("repetition_penalty", 1.1)),
+            "temperature": float(self.config.get("temperature", 0.8)),
+            "cache_reference": bool(self.config.get("cache_reference", True)),
+            "context_length": int(self.config.get("context_length", 8192)),
+        }
+        cache_key = self.audio_cache.generate_cache_key(
+            "fish_audio_s2", text=text, audio_component="|".join(components),
+            reference_text="|".join(reference_texts), character=cache_character, **params,
+        ) if enable_audio_cache else None
+        if cache_key:
+            cached = self.audio_cache.get_cached_audio(cache_key)
+            if cached:
+                print("💾 Fish Audio S2: Using cached audio")
+                return cached[0]
+        audio, sample_rate = self._engine().generate(text=text, references=references, **params)
+        if sample_rate != self.SAMPLE_RATE:
+            raise RuntimeError(f"Fish S2 returned unexpected sample rate {sample_rate}")
+        audio = audio.detach().float().cpu()
+        if audio.dim() == 1:
+            audio = audio.unsqueeze(0)
+        if cache_key:
+            self.audio_cache.cache_audio(cache_key, audio, audio.shape[-1] / self.SAMPLE_RATE)
+        return audio

engines/adapters/higgs_audio_adapter.py

@@ -14,6 +14,8 @@ if parent_dir not in sys.path:
     sys.path.insert(0, parent_dir)
 
 from utils.models.language_mapper import get_model_for_language
+from utils.models.engine_registry import get_default_runtime_profile
+from utils.models.factory_config import normalize_runtime_mode, runtime_uses_isolation
 from engines.higgs_audio.higgs_audio import HiggsAudioEngine
 from engines.higgs_audio.higgs_audio_downloader import HIGGS_AUDIO_MODELS
 
@@ -31,6 +33,9 @@ class HiggsAudioEngineAdapter:
         self.node = node_instance
         self.engine_type = "higgs_audio"
         self.higgs_engine = HiggsAudioEngine()
+        node_config = getattr(node_instance, "config", {}) or {}
+        self._runtime_mode = normalize_runtime_mode(node_config.get("runtime_mode"))
+        self._runtime_profile = node_config.get("runtime_profile")
     
     def get_model_for_language(self, lang_code: str, default_model: str) -> str:
         """
@@ -85,6 +90,23 @@ class HiggsAudioEngineAdapter:
             # Default fallback
             generation_model = "bosonai/higgs-audio-v2-generation-3B-base"
             tokenizer_model = "bosonai/higgs-audio-v2-tokenizer"
+
+        if runtime_uses_isolation(self._runtime_mode):
+            from utils.models.unified_model_interface import load_tts_model
+
+            self.higgs_engine = load_tts_model(
+                engine_name="higgs_audio",
+                model_name=model_name,
+                device=device,
+                model_path=generation_model,
+                tokenizer_path=tokenizer_model,
+                enable_cuda_graphs=enable_cuda_graphs,
+                runtime_mode=self._runtime_mode,
+                runtime_profile=self._runtime_profile or get_default_runtime_profile("higgs_audio"),
+            )
+            self.node.current_model_name = model_name
+            print(f"✅ Higgs Audio adapter: Loaded model '{model_name}' on {device} via shared runtime")
+            return
         
         # Initialize the Higgs Audio engine with CUDA Graph setting
         self.higgs_engine.initialize_engine(
@@ -138,7 +160,7 @@ class HiggsAudioEngineAdapter:
         top_p = params.get("top_p", 0.95)
         top_k = params.get("top_k", 50)
         max_new_tokens = params.get("max_new_tokens", 2048)
-        force_audio_gen = params.get("force_audio_gen", False)
+        force_audio_gen = params.get("force_audio_gen", True)
         ras_win_len = params.get("ras_win_len", 7)  # Default from boson_multimodal
         ras_max_num_repeat = params.get("ras_max_num_repeat", 2)  # Default from boson_multimodal
         enable_cuda_graphs = params.get("enable_cuda_graphs", True)  # CUDA Graph toggle
@@ -161,7 +183,7 @@ class HiggsAudioEngineAdapter:
         max_tokens_per_chunk = int(max_chars_per_chunk / 3.5)
         
         # Initialize engine if not already done
-        if not self.higgs_engine.engine:
+        if not getattr(self.higgs_engine, "engine", None):
             print(f"🚀 Initializing Higgs Audio engine with model: {model_name}")
             self.load_base_model(model_name, device)
         
@@ -394,7 +416,7 @@ class HiggsAudioEngineAdapter:
         validated["max_new_tokens"] = max(1, min(4096, int(max_tokens)))
         
         # Force audio gen validation
-        force_audio_gen = params.get("force_audio_gen", False)
+        force_audio_gen = params.get("force_audio_gen", True)
         validated["force_audio_gen"] = bool(force_audio_gen)
         
         # RAS parameters validation
@@ -426,4 +448,4 @@ class HiggsAudioEngineAdapter:
         # Cache settings
         validated["enable_audio_cache"] = params.get("enable_audio_cache", True)
         
-        return validated
+        return validated

engines/adapters/higgs_audio_v3_adapter.py

@@ -0,0 +1,206 @@
+"""Higgs Audio v3 adapter for the unified TTS pipeline."""
+
+from __future__ import annotations
+
+import os
+import sys
+from typing import Any, Dict, Optional, Tuple
+
+import torch
+
+current_dir = os.path.dirname(__file__)
+engines_dir = os.path.dirname(current_dir)
+project_root = os.path.dirname(engines_dir)
+if project_root not in sys.path:
+    sys.path.insert(0, project_root)
+
+from utils.audio.audio_hash import generate_stable_audio_component
+from utils.audio.cache import get_audio_cache
+from utils.models.factory_config import ModelLoadConfig
+from utils.text.pause_processor import PauseTagProcessor
+
+
+class HiggsAudioV3EngineAdapter:
+    """Adapter for native Higgs Audio v3 generation."""
+
+    SAMPLE_RATE = 24000
+
+    def __init__(self, node_instance=None):
+        self.node = node_instance
+        self.audio_cache = get_audio_cache()
+        self._last_config: Optional[ModelLoadConfig] = None
+
+    def load_model(
+        self,
+        model: str,
+        device: str = "auto",
+        dtype: str = "auto",
+        attention: str = "auto",
+    ):
+        from utils.models.unified_model_interface import unified_model_interface
+
+        config = ModelLoadConfig(
+            engine_name="higgs_audio_v3",
+            model_type="tts",
+            model_name=model,
+            model_path=model,
+            device=device,
+            additional_params={
+                "dtype": dtype,
+                "attention": attention,
+            },
+        )
+        self._last_config = config
+        unified_model_interface.load_model(config)
+
+    def update_model_config(
+        self,
+        model: str,
+        device: str = "auto",
+        dtype: str = "auto",
+        attention: str = "auto",
+    ):
+        self.load_model(model, device, dtype, attention)
+
+    def _get_engine(self):
+        if self._last_config is None:
+            raise RuntimeError("Higgs Audio v3 model has not been loaded")
+        from utils.models.unified_model_interface import unified_model_interface
+
+        return unified_model_interface.load_model(self._last_config)
+
+    def generate_with_pause_tags(
+        self,
+        text: str,
+        voice_ref: Optional[Dict[str, Any]],
+        params: Dict[str, Any],
+        process_pauses: bool = True,
+        character_name: Optional[str] = None,
+    ) -> torch.Tensor:
+        engine = self._get_engine()
+        if process_pauses and PauseTagProcessor.has_pause_tags(text):
+            return self._generate_with_pauses(text, voice_ref, params, character_name, engine)
+        return self._generate_direct(text, voice_ref, params, character_name, engine)
+
+    def _generate_direct(
+        self,
+        text: str,
+        voice_ref: Optional[Dict[str, Any]],
+        params: Dict[str, Any],
+        character_name: Optional[str] = None,
+        engine=None,
+    ) -> torch.Tensor:
+        if engine is None:
+            engine = self._get_engine()
+
+        reference_audio, reference_text, audio_component = self._extract_voice_reference(voice_ref)
+        model_name = params.get("model", "higgs-audio-v3-tts-4b")
+        seed = int(params.get("seed", 0) or 0)
+        max_new_tokens = int(params.get("max_new_tokens", 2048))
+        temperature = float(params.get("temperature", 1.0))
+        top_p = float(params.get("top_p", 0.95))
+        top_k = int(params.get("top_k", 50))
+
+        cache_key = self.audio_cache.generate_cache_key(
+            "higgs_audio_v3",
+            text=text,
+            model_variant=model_name,
+            audio_component=audio_component,
+            reference_text=reference_text,
+            temperature=temperature,
+            top_p=top_p,
+            top_k=top_k,
+            max_new_tokens=max_new_tokens,
+            seed=seed,
+            device=params.get("device", "auto"),
+            dtype=params.get("dtype", "auto"),
+            attention=params.get("attention", "auto"),
+            character=character_name or "narrator",
+        )
+
+        if params.get("enable_audio_cache", True):
+            cached_audio = self.audio_cache.get_cached_audio(cache_key)
+            if cached_audio:
+                print(f"💾 Using cached Higgs Audio v3 audio for '{character_name or 'narrator'}': '{text[:30]}...'")
+                return cached_audio[0]
+
+        audio_tensor, sample_rate = engine.generate(
+            text=text,
+            reference_audio=reference_audio,
+            reference_text=reference_text,
+            max_new_tokens=max_new_tokens,
+            temperature=temperature,
+            top_p=top_p,
+            top_k=top_k,
+            seed=seed,
+        )
+
+        if sample_rate != self.SAMPLE_RATE:
+            raise RuntimeError(f"Higgs Audio v3 returned {sample_rate} Hz; expected {self.SAMPLE_RATE} Hz")
+
+        if params.get("enable_audio_cache", True):
+            duration = self.audio_cache._calculate_duration(audio_tensor, "higgs_audio_v3")
+            self.audio_cache.cache_audio(cache_key, audio_tensor, duration)
+        return audio_tensor
+
+    def _generate_with_pauses(
+        self,
+        text: str,
+        voice_ref: Optional[Dict[str, Any]],
+        params: Dict[str, Any],
+        character_name: Optional[str] = None,
+        engine=None,
+    ) -> torch.Tensor:
+        segments, _ = PauseTagProcessor.parse_pause_tags(text)
+        audio_parts = []
+        for segment_type, content in segments:
+            if segment_type == "text":
+                audio_tensor = self._generate_direct(content, voice_ref, params, character_name, engine)
+                if audio_tensor.dim() == 1:
+                    audio_tensor = audio_tensor.unsqueeze(0)
+                elif audio_tensor.dim() == 3:
+                    audio_tensor = audio_tensor.squeeze(0)
+                audio_parts.append(audio_tensor)
+            elif segment_type == "pause":
+                audio_parts.append(PauseTagProcessor.create_silence_segment(content, self.SAMPLE_RATE))
+
+        if not audio_parts:
+            return torch.zeros(1, 0)
+        return torch.cat(audio_parts, dim=-1)
+
+    def _extract_voice_reference(self, voice_ref: Optional[Dict[str, Any]]) -> Tuple[Optional[Dict[str, Any]], str, str]:
+        if not voice_ref or not isinstance(voice_ref, dict):
+            return None, "", "zero_shot"
+
+        ref_audio = None
+        for key in ("audio", "audio_dict", "reference_audio", "waveform"):
+            if key in voice_ref and voice_ref.get(key) is not None:
+                ref_audio = voice_ref.get(key)
+                break
+        reference_text = (
+            voice_ref.get("reference_text")
+            or voice_ref.get("text")
+            or voice_ref.get("prompt_text")
+            or ""
+        )
+
+        if ref_audio is None and "audio_path" in voice_ref:
+            audio_path = voice_ref.get("audio_path")
+            if audio_path and os.path.exists(audio_path):
+                try:
+                    from utils.audio.processing import AudioProcessingUtils
+
+                    waveform, sample_rate = AudioProcessingUtils.safe_load_audio(audio_path)
+                    ref_audio = {"waveform": waveform, "sample_rate": sample_rate}
+                except Exception as e:
+                    print(f"⚠️ Failed to load Higgs Audio v3 reference audio {audio_path}: {e}")
+
+        if isinstance(ref_audio, dict) and "waveform" in ref_audio:
+            return ref_audio, str(reference_text), generate_stable_audio_component(reference_audio=ref_audio)
+
+        if torch.is_tensor(ref_audio):
+            sample_rate = int(voice_ref.get("sample_rate", self.SAMPLE_RATE))
+            audio_dict = {"waveform": ref_audio, "sample_rate": sample_rate}
+            return audio_dict, str(reference_text), generate_stable_audio_component(reference_audio=audio_dict)
+
+        return None, str(reference_text), "zero_shot"

engines/adapters/index_tts_adapter.py

@@ -12,7 +12,8 @@ from typing import Dict, Any, Optional, List, Union
 from engines.index_tts.index_tts import IndexTTSEngine
 from engines.index_tts.index_tts_downloader import index_tts_downloader
 from utils.text.character_parser import character_parser
-from utils.voice.discovery import get_character_mapping, get_available_characters
+from utils.voice.discovery import get_character_mapping, get_available_characters
+from utils.voice.character_logging import resolved_character_label
 from utils.audio.cache import get_audio_cache
 
 
@@ -76,7 +77,26 @@ class IndexTTSAdapter:
         )
         
     
-    def generate(self,
+    @staticmethod
+    def _normalize_emotion_audio(emotion_audio):
+        """Return a path or waveform dict accepted by IndexTTS-2."""
+        if not isinstance(emotion_audio, dict):
+            return emotion_audio
+        if emotion_audio.get("audio_path"):
+            print(
+                "🎭 Using Character Voices emotion audio: "
+                f"{emotion_audio.get('character_name', 'unknown')} -> "
+                f"{emotion_audio['audio_path']}"
+            )
+            return emotion_audio["audio_path"]
+        if "waveform" in emotion_audio:
+            return emotion_audio
+        nested_audio = emotion_audio.get("audio")
+        if isinstance(nested_audio, dict) and "waveform" in nested_audio:
+            return nested_audio
+        return emotion_audio
+
+    def generate(self,
                 text: str,
                 speaker_audio: Optional[str] = None,
                 emotion_audio: Optional[str] = None,
@@ -93,9 +113,12 @@ class IndexTTSAdapter:
                 top_k: int = 30,
                 length_penalty: float = 0.0,
                 num_beams: int = 3,
-                repetition_penalty: float = 10.0,
-                max_mel_tokens: int = 1500,
-                # Streaming parameters
+                repetition_penalty: float = 10.0,
+                max_mel_tokens: int = 1500,
+                language: str = "English",
+                duration_factor: float = 1.0,
+                text_normalization: bool = True,
+                # Streaming parameters
                 stream_return: bool = False,
                 more_segment_before: int = 0,
                 **kwargs) -> torch.Tensor:
@@ -119,7 +142,10 @@ class IndexTTSAdapter:
             length_penalty: Length penalty for beam search
             num_beams: Number of beams for beam search
             repetition_penalty: Repetition penalty
-            max_mel_tokens: Maximum mel tokens to generate
+            max_mel_tokens: Maximum mel tokens to generate
+            language: IndexTTS-2.5 language code/name
+            duration_factor: Official 2.5 internal feature-duration multiplier
+            text_normalization: Enable multilingual text normalization
             **kwargs: Additional parameters
             
         Returns:
@@ -135,9 +161,31 @@ class IndexTTSAdapter:
             # Parse character switching tags with emotion support
             processed_segments = self._process_character_tags_with_emotions(text)
             
-            if len(processed_segments) > 1:
-                # Multi-segment character switching - process each segment separately
-                return self._generate_multi_character_segments(processed_segments, speaker_audio, emotion_audio, **kwargs)
+            if len(processed_segments) > 1:
+                # Multi-segment character switching - process each segment separately
+                return self._generate_multi_character_segments(
+                    processed_segments, speaker_audio, emotion_audio,
+                    emotion_alpha=emotion_alpha,
+                    emotion_vector=emotion_vector,
+                    use_emotion_text=use_emotion_text,
+                    emotion_text=emotion_text,
+                    use_random=use_random,
+                    interval_silence=interval_silence,
+                    max_text_tokens_per_segment=max_text_tokens_per_segment,
+                    temperature=temperature,
+                    top_p=top_p,
+                    top_k=top_k,
+                    length_penalty=length_penalty,
+                    num_beams=num_beams,
+                    repetition_penalty=repetition_penalty,
+                    max_mel_tokens=max_mel_tokens,
+                    language=language,
+                    duration_factor=duration_factor,
+                    text_normalization=text_normalization,
+                    stream_return=stream_return,
+                    more_segment_before=more_segment_before,
+                    **kwargs,
+                )
             elif processed_segments:
                 # Single character segment
                 first_segment = processed_segments[0]
@@ -157,19 +205,10 @@ class IndexTTSAdapter:
         # Determine final speaker and emotion audio
         final_speaker_audio = speaker_audio
 
-        # Handle Character Voices emotion_audio format
-        if emotion_audio and isinstance(emotion_audio, dict):
-            if "audio_path" in emotion_audio:
-                # Character Voices format: {'audio': {...}, 'audio_path': 'path', ...}
-                final_emotion_audio = emotion_audio["audio_path"]
-                print(f"🎭 Using Character Voices emotion audio: {emotion_audio.get('character_name', 'unknown')} -> {final_emotion_audio}")
-            elif "waveform" in emotion_audio:
-                # Direct AUDIO format: {'waveform': tensor, 'sample_rate': rate}
-                final_emotion_audio = emotion_audio
-            else:
-                final_emotion_audio = emotion_audio
-        else:
-            final_emotion_audio = emotion_audio
+        # Normalize the two supported audio-reference shapes:
+        # Character Voices returns {audio: {waveform, sample_rate}, audio_path: ...},
+        # while ComfyUI AUDIO returns {waveform, sample_rate} directly.
+        final_emotion_audio = self._normalize_emotion_audio(emotion_audio)
         
         # Only do character mapping if we actually have character tags
         if has_character_tags:
@@ -221,9 +260,12 @@ class IndexTTSAdapter:
             length_penalty=length_penalty,
             repetition_penalty=repetition_penalty,
             max_mel_tokens=max_mel_tokens,
-            max_text_tokens_per_segment=max_text_tokens_per_segment,
-            interval_silence=interval_silence,
-            stream_return=stream_return,
+            max_text_tokens_per_segment=max_text_tokens_per_segment,
+            interval_silence=interval_silence,
+            language=language,
+            duration_factor=duration_factor,
+            text_normalization=text_normalization,
+            stream_return=stream_return,
             more_segment_before=more_segment_before,
             **kwargs  # Include seed and other kwargs in cache key
         )
@@ -263,18 +305,11 @@ class IndexTTSAdapter:
         engine_kwargs['stream_return'] = stream_return
         engine_kwargs['more_segment_before'] = more_segment_before
 
-        # Apply consistent emotion priority: emotion_audio takes precedence over other emotion controls
-        # This ensures consistent behavior whether using character tags or direct engine inputs
-        if final_emotion_audio:
-            # emotion_audio connected - disable other emotion controls
-            final_emotion_vector = None
-            final_use_emotion_text = False
-            final_emotion_text = None
-        else:
-            # No emotion_audio - use provided emotion controls
-            final_emotion_vector = emotion_vector
-            final_use_emotion_text = use_emotion_text
-            final_emotion_text = emotion_text
+        # Audio emotion and vector/text emotion are independent conditioning
+        # sources.  IndexTTS-2 blends them in its latent emotion space.
+        final_emotion_vector = emotion_vector
+        final_use_emotion_text = use_emotion_text
+        final_emotion_text = emotion_text
 
         # Generate audio with OOM protection
         try:
@@ -294,9 +329,12 @@ class IndexTTSAdapter:
                 top_k=top_k,
                 length_penalty=length_penalty,
                 num_beams=num_beams,
-                repetition_penalty=repetition_penalty,
-                max_mel_tokens=max_mel_tokens,
-                **engine_kwargs
+                repetition_penalty=repetition_penalty,
+                max_mel_tokens=max_mel_tokens,
+                language=language,
+                duration_factor=duration_factor,
+                text_normalization=text_normalization,
+                **engine_kwargs
             )
         except torch.OutOfMemoryError as e:
             # Analyze audio after OOM to provide helpful feedback
@@ -376,7 +414,7 @@ class IndexTTSAdapter:
         Returns:
             Combined audio tensor [1, samples] at 22050 Hz
         """
-        audio_segments = []
+        audio_segments = []
         
         # Get character mapping for all unique characters
         unique_characters = set()
@@ -391,12 +429,22 @@ class IndexTTSAdapter:
         if unique_characters:
             character_mapping = get_character_mapping(list(unique_characters), engine_type="index_tts")
         
-        print(f"🎭 IndexTTS-2: Processing {len(segments)} character segment(s) - {', '.join([s.get('character', 'narrator') for s in segments])}")
+        resolved_names = [
+            resolved_character_label(
+                segment.get('character', 'narrator'),
+                character_mapping.get(segment.get('character', 'narrator'), (default_speaker_audio, None)),
+            )
+            for segment in segments
+        ]
+        print(f"🎭 IndexTTS-2: Processing {len(segments)} character segment(s) - {', '.join(resolved_names)}")
         
         for segment in segments:
             character_name = segment.get('character', 'narrator')
             segment_text = segment.get('text', '').strip()
-            emotion_ref = segment.get('emotion')
+            emotion_ref = segment.get('emotion')
+            segment_kwargs = dict(kwargs)
+            if segment.get('language'):
+                segment_kwargs['language'] = segment['language']
             
             if not segment_text:
                 continue
@@ -407,7 +455,7 @@ class IndexTTSAdapter:
                 character_audio_path = character_mapping[character_name][0]
                 if character_audio_path:
                     speaker_audio = character_audio_path
-                    print(f"📖 Using character voice '{character_name}' | Ref: '{speaker_audio}'")
+                    print(f"📖 Using character voice '{resolved_character_label(character_name, speaker_audio)}' | Ref: '{speaker_audio}'")
                 else:
                     print(f"⚠️ Character '{character_name}' has no audio reference, using default")
             
@@ -422,24 +470,24 @@ class IndexTTSAdapter:
             # Generate cache key for this segment
             segment_cache_key = self._generate_cache_key(
                 text=segment_text,
-                speaker_audio=speaker_audio,
-                emotion_audio=emotion_audio,
-                **kwargs
+                speaker_audio=speaker_audio,
+                emotion_audio=emotion_audio,
+                **segment_kwargs
             )
 
             # Check cache first
             cached_segment_audio = self.audio_cache.get_cached_audio(segment_cache_key)
             if cached_segment_audio:
-                print(f"💾 Using cached IndexTTS-2 segment for '{character_name}': '{segment_text[:30]}...'")
+                print(f"💾 Using cached IndexTTS-2 segment for '{resolved_character_label(character_name, speaker_audio)}': '{segment_text[:30]}...'")
                 segment_audio = cached_segment_audio[0]
             else:
                 # Generate audio for this segment with OOM protection
                 try:
                     segment_audio = self.engine.generate(
                         text=segment_text,
-                        speaker_audio=speaker_audio,
-                        emotion_audio=emotion_audio,
-                        **kwargs
+                        speaker_audio=speaker_audio,
+                        emotion_audio=emotion_audio,
+                        **segment_kwargs
                     )
                 except torch.OutOfMemoryError as e:
                     # Analyze audio after OOM in multi-character segments
@@ -463,9 +511,18 @@ class IndexTTSAdapter:
             # Return silence if no segments generated
             return torch.zeros(1, 22050, dtype=torch.float32)
     
-    def _generate_cache_key(self, **params) -> str:
-        """Generate cache key for IndexTTS-2."""
-        return self.audio_cache.generate_cache_key('index_tts', **params)
+    def _generate_cache_key(self, **params) -> str:
+        """Generate cache key for IndexTTS-2."""
+        model_identity = {}
+        if self.engine is not None:
+            model_identity = {
+                "model_name": getattr(self.engine, "model_name", None),
+                "model_version": getattr(self.engine, "model_version", None),
+                "model_path": getattr(self.engine, "model_dir", None),
+            }
+        return self.audio_cache.generate_cache_key(
+            'index_tts', **model_identity, **params
+        )
 
     def _analyze_audio_after_oom(self, speaker_audio: str, emotion_audio: str, max_mel_tokens: int) -> str:
         """
@@ -534,16 +591,18 @@ class IndexTTSAdapter:
 
         return "\n".join(analysis_parts)
 
-    def _get_stable_audio_identifier(self, audio_path: str) -> str:
-        """
-        Get stable identifier for audio file using centralized audio hashing.
-        """
-        if not audio_path:
-            return audio_path
-
-        # Use our centralized audio hashing utility
-        from utils.audio.audio_hash import generate_stable_audio_component
-        return generate_stable_audio_component(audio_file_path=audio_path)
+    def _get_stable_audio_identifier(self, audio_path: str) -> str:
+        """
+        Get stable identifier for audio file using centralized audio hashing.
+        """
+        if not audio_path:
+            return audio_path
+
+        # Use our centralized audio hashing utility
+        from utils.audio.audio_hash import generate_stable_audio_component
+        if isinstance(audio_path, dict):
+            return generate_stable_audio_component(reference_audio=audio_path)
+        return generate_stable_audio_component(audio_file_path=audio_path)
     
     def get_supported_formats(self) -> List[str]:
         """Get supported audio formats."""
@@ -580,10 +639,6 @@ class IndexTTSAdapter:
     
     def unload(self):
         """Unload the engine to free memory."""
-        if self.engine:
-            self.engine.unload()
-            self.engine = None
-    
-    def __del__(self):
-        """Cleanup on deletion."""
-        self.unload()
+        if self.engine:
+            self.engine.unload()
+            self.engine = None

engines/adapters/moss_soundeffect_v2_adapter.py

@@ -0,0 +1,69 @@
+"""Suite adapter for MOSS-SoundEffect v2."""
+
+from typing import Any, Dict, Tuple
+
+import torch
+
+from engines.moss_soundeffect_v2.downloader import MossSoundEffectV2Downloader
+from utils.audio.cache import get_audio_cache
+from utils.models.factory_config import ModelLoadConfig
+from utils.models.unified_model_interface import unified_model_interface
+
+
+class MossSoundEffectV2Adapter:
+    def __init__(self, config: Dict[str, Any]):
+        self.config = dict(config)
+        self.audio_cache = get_audio_cache()
+        self._load_config = None
+
+    def _load(self):
+        model = self.config.get("model", MossSoundEffectV2Downloader.MODEL_NAME)
+        model_path = MossSoundEffectV2Downloader().resolve_model_path(model)
+        self._load_config = ModelLoadConfig(
+            engine_name="moss_soundeffect_v2",
+            model_type="tts",
+            model_name=str(model).removeprefix("local:"),
+            model_path=model_path,
+            device=self.config.get("device", "auto"),
+            runtime_mode="main_environment",
+            additional_params={"dtype": self.config.get("dtype", "auto")},
+        )
+        return unified_model_interface.load_model(self._load_config)
+
+    def generate(
+        self,
+        description: str,
+        duration_seconds: float,
+        seed: int,
+        enable_audio_cache: bool,
+    ) -> Tuple[torch.Tensor, int, bool]:
+        duration_seconds = round(float(duration_seconds), 1)
+        if duration_seconds > 30.0:
+            raise ValueError(
+                "MOSS-SoundEffect v2 supports a maximum duration of 30 seconds. "
+                "Lower duration_seconds in the 🌩️ Sound Effects node."
+            )
+        params = {
+            "description": description,
+            "model": self.config.get("model"),
+            "duration_seconds": duration_seconds,
+            "inference_steps": self.config.get("inference_steps", 100),
+            "cfg_scale": self.config.get("cfg_scale", 4.0),
+            "sigma_shift": self.config.get("sigma_shift", 5.0),
+            "negative_prompt": self.config.get("negative_prompt", ""),
+            "seed": int(seed),
+            "dtype": self.config.get("dtype", "auto"),
+            "device": self.config.get("device", "auto"),
+        }
+        cache_key = self.audio_cache.generate_cache_key("moss_soundeffect_v2", **params)
+        if enable_audio_cache:
+            cached = self.audio_cache.get_cached_audio(cache_key)
+            if cached is not None:
+                return cached[0], 48000, True
+
+        engine = self._load()
+        waveform, sample_rate = engine.generate_sound_effect(**params)
+        waveform = waveform.detach().cpu().float()
+        if enable_audio_cache:
+            self.audio_cache.cache_audio(cache_key, waveform, waveform.shape[-1] / float(sample_rate))
+        return waveform, int(sample_rate), False

engines/adapters/moss_tts_adapter.py

@@ -21,6 +21,7 @@ from utils.audio.audio_hash import generate_stable_audio_component
 from utils.audio.cache import get_audio_cache
 from utils.models.factory_config import ModelLoadConfig
 from utils.text.pause_processor import PauseTagProcessor
+from utils.voice.reference import effective_voice_audio
 
 
 class MossTTSEngineAdapter:
@@ -40,6 +41,7 @@ class MossTTSEngineAdapter:
         attn_implementation: str = "auto",
         codec_model: str = "MOSS-Audio-Tokenizer",
         lora_adapter: Optional[str] = None,
+        defer_load: bool = False,
     ):
         from utils.models.unified_model_interface import unified_model_interface
 
@@ -57,7 +59,8 @@ class MossTTSEngineAdapter:
             },
         )
         self._last_config = config
-        unified_model_interface.load_model(config)
+        if not defer_load:
+            unified_model_interface.load_model(config)
 
     def update_model_config(
         self,
@@ -104,9 +107,6 @@ class MossTTSEngineAdapter:
         character_name: Optional[str] = None,
         engine=None,
     ) -> torch.Tensor:
-        if engine is None:
-            engine = self._get_engine()
-
         reference_audio, reference_sample_rate, audio_component = self._extract_voice_reference(voice_ref)
         model_variant = params.get("model_variant", "MOSS-TTS-Local-Transformer")
         language = params.get("language", "auto")
@@ -145,11 +145,15 @@ class MossTTSEngineAdapter:
             character=character_name or "narrator",
         )
 
-        cached_audio = self.audio_cache.get_cached_audio(cache_key)
+        enable_audio_cache = bool(params.get("enable_audio_cache", True))
+        cached_audio = self.audio_cache.get_cached_audio(cache_key) if enable_audio_cache else None
         if cached_audio:
             print(f"💾 Using cached MOSS-TTS audio for '{character_name or 'narrator'}': '{text[:30]}...'")
             return cached_audio[0]
 
+        if engine is None:
+            engine = self._get_engine()
+
         audio_tensor, sample_rate = engine.generate(
             text=text,
             reference_audio=reference_audio,
@@ -174,10 +178,24 @@ class MossTTSEngineAdapter:
         if sample_rate != self.SAMPLE_RATE:
             raise RuntimeError(f"MOSS-TTS returned unexpected sample rate {sample_rate}; expected {self.SAMPLE_RATE}")
 
-        duration = self.audio_cache._calculate_duration(audio_tensor, "moss_tts")
-        self.audio_cache.cache_audio(cache_key, audio_tensor, duration)
+        if enable_audio_cache:
+            duration = self.audio_cache._calculate_duration(audio_tensor, "moss_tts")
+            self.audio_cache.cache_audio(cache_key, audio_tensor, duration)
         return audio_tensor
 
+    def generate_sound_effect(self, description: str, params: Dict[str, Any]) -> torch.Tensor:
+        """Generate MOSS-SoundEffect v1 audio without a speech or narrator reference."""
+        sound_params = dict(params)
+        sound_params["ambient_sound"] = str(description or "").strip()
+        if not sound_params["ambient_sound"]:
+            raise ValueError("MOSS-SoundEffect requires a non-empty description")
+        return self._generate_direct(
+            text="",
+            voice_ref=None,
+            params=sound_params,
+            character_name="sound_effects",
+        )
+
     def _generate_with_pauses(
         self,
         text: str,
@@ -207,12 +225,7 @@ class MossTTSEngineAdapter:
         if not voice_ref or not isinstance(voice_ref, dict):
             return None, None, "default_voice"
 
-        ref_audio = (
-            voice_ref.get("audio_path")
-            or voice_ref.get("prompt_audio_path")
-            or voice_ref.get("audio")
-            or voice_ref.get("waveform")
-        )
+        ref_audio = effective_voice_audio(voice_ref)
         if ref_audio is None:
             return None, None, "default_voice"
 
@@ -248,12 +261,7 @@ class MossTTSEngineAdapter:
             print(f"❌ MOSS-TTSD {speaker_label}: invalid voice reference type {type(voice_ref).__name__}")
             return None, "invalid_voice"
 
-        ref_audio = (
-            voice_ref.get("audio_path")
-            or voice_ref.get("prompt_audio_path")
-            or voice_ref.get("audio")
-            or voice_ref.get("waveform")
-        )
+        ref_audio = effective_voice_audio(voice_ref)
         reference_text = (
             voice_ref.get("reference_text")
             or voice_ref.get("text")

engines/adapters/omnivoice_adapter.py

@@ -0,0 +1,299 @@
+"""
+OmniVoice engine adapter.
+
+Translates TTS Audio Suite processor calls into the official OmniVoice runtime.
+"""
+
+import os
+import sys
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+
+current_dir = os.path.dirname(__file__)
+engines_dir = os.path.dirname(current_dir)
+project_root = os.path.dirname(engines_dir)
+if project_root not in sys.path:
+    sys.path.insert(0, project_root)
+
+from utils.audio.audio_hash import generate_stable_audio_component
+from utils.audio.cache import get_audio_cache
+from utils.models.factory_config import ModelLoadConfig
+from utils.models.language_mapper import resolve_language_alias
+from utils.voice.reference import effective_voice_audio
+
+
+class OmniVoiceEngineAdapter:
+    """Adapter for official OmniVoice inference."""
+
+    SAMPLE_RATE = 24000
+
+    def __init__(self, config: Optional[Dict[str, Any]] = None):
+        self.config = config.copy() if config else {}
+        self.audio_cache = get_audio_cache()
+        self._last_config: Optional[ModelLoadConfig] = None
+        self._load_signature = None
+
+    def update_config(self, new_config: Dict[str, Any]):
+        self.config = new_config.copy() if new_config else {}
+
+    def _build_load_signature(self) -> Tuple[Any, ...]:
+        return (
+            self.config.get("model_variant", "OmniVoice"),
+            self.config.get("device", "auto"),
+            self.config.get("dtype", "auto"),
+        )
+
+    def load_model(
+        self,
+        model_variant: str,
+        device: str = "auto",
+        dtype: str = "auto",
+    ):
+        from utils.models.unified_model_interface import unified_model_interface
+
+        config = ModelLoadConfig(
+            engine_name="omnivoice",
+            model_type="tts",
+            model_name=model_variant,
+            model_path=model_variant,
+            device=device,
+            additional_params={
+                "dtype": dtype,
+            },
+        )
+        self._last_config = config
+        unified_model_interface.load_model(config)
+
+    def _ensure_model_loaded(self):
+        signature = self._build_load_signature()
+        if signature == self._load_signature and self._last_config is not None:
+            return
+
+        self.load_model(
+            model_variant=self.config.get("model_variant", "OmniVoice"),
+            device=self.config.get("device", "auto"),
+            dtype=self.config.get("dtype", "auto"),
+        )
+        self._load_signature = signature
+
+    def _get_engine(self):
+        if self._last_config is None:
+            self._ensure_model_loaded()
+        from utils.models.unified_model_interface import unified_model_interface
+
+        return unified_model_interface.load_model(self._last_config)
+
+    @staticmethod
+    def _seed_everything(seed: int):
+        if int(seed or 0) <= 0:
+            return
+        torch.manual_seed(int(seed))
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed_all(int(seed))
+        if hasattr(torch, "xpu") and torch.xpu.is_available():
+            torch.xpu.manual_seed_all(int(seed))
+
+    @staticmethod
+    def _first_non_none(*values):
+        for value in values:
+            if value is None:
+                continue
+            if isinstance(value, str) and not value.strip():
+                continue
+            return value
+        return None
+
+    @staticmethod
+    def _normalize_waveform_tensor(waveform: torch.Tensor) -> torch.Tensor:
+        if waveform.dim() == 3:
+            waveform = waveform.squeeze(0)
+        if waveform.dim() == 1:
+            waveform = waveform.unsqueeze(0)
+        return waveform.detach().float().cpu()
+
+    def _extract_voice_reference(
+        self,
+        voice_ref: Optional[Dict[str, Any]],
+    ) -> Tuple[Optional[Union[str, Tuple[torch.Tensor, int]]], str, str]:
+        if not voice_ref or not isinstance(voice_ref, dict):
+            return None, "", "default_voice"
+
+        prompt_text = (
+            voice_ref.get("reference_text")
+            or voice_ref.get("prompt_text")
+            or voice_ref.get("text")
+            or ""
+        ).strip()
+
+        ref_audio = effective_voice_audio(voice_ref)
+
+        if ref_audio is None:
+            return None, prompt_text, "default_voice"
+
+        if isinstance(ref_audio, str):
+            return ref_audio, prompt_text, generate_stable_audio_component(audio_file_path=ref_audio)
+
+        if isinstance(ref_audio, dict) and "waveform" in ref_audio:
+            waveform = self._normalize_waveform_tensor(ref_audio["waveform"])
+            sample_rate = int(ref_audio.get("sample_rate", self.SAMPLE_RATE))
+            audio_dict = {"waveform": waveform, "sample_rate": sample_rate}
+            return (
+                (waveform, sample_rate),
+                prompt_text,
+                generate_stable_audio_component(reference_audio=audio_dict),
+            )
+
+        if torch.is_tensor(ref_audio):
+            waveform = self._normalize_waveform_tensor(ref_audio)
+            sample_rate = int(voice_ref.get("sample_rate", self.SAMPLE_RATE))
+            audio_dict = {"waveform": waveform, "sample_rate": sample_rate}
+            return (
+                (waveform, sample_rate),
+                prompt_text,
+                generate_stable_audio_component(reference_audio=audio_dict),
+            )
+
+        raise TypeError(f"Unsupported OmniVoice reference type: {type(ref_audio)}")
+
+    @staticmethod
+    def _normalize_language(language: Optional[str]) -> Optional[str]:
+        if language is None:
+            return None
+
+        normalized = str(language).strip()
+        if not normalized:
+            return None
+
+        lowered = normalized.lower()
+        if lowered in {"auto", "none"}:
+            return None
+
+        resolved = resolve_language_alias(normalized)
+
+        # OmniVoice uses the base ISO code for Portuguese rather than regional
+        # variants accepted by other suite engines.
+        if resolved in {"pt", "pt-br", "pt-pt"}:
+            return "pt"
+
+        if resolved and resolved.lower() != lowered:
+            return resolved
+        return normalized
+
+    def generate_single(
+        self,
+        text: str,
+        voice_ref: Optional[Dict[str, Any]],
+        seed: int = 0,
+        enable_audio_cache: bool = True,
+        character_name: Optional[str] = None,
+    ) -> torch.Tensor:
+        stripped = (text or "").strip()
+        if not stripped:
+            return torch.zeros(1, 0, dtype=torch.float32)
+
+        self._ensure_model_loaded()
+        engine = self._get_engine()
+
+        ref_audio, prompt_text, audio_component = self._extract_voice_reference(voice_ref)
+        language = self._normalize_language(self.config.get("language", "Auto"))
+        instruct = str(self.config.get("instruct", "") or "").strip() or None
+        model_variant = self.config.get("model_variant", "OmniVoice")
+        device = self.config.get("device", "auto")
+        dtype = self.config.get("dtype", "auto")
+        num_step = int(self.config.get("num_steps", 32))
+        guidance_scale = float(self.config.get("guidance_scale", 2.0))
+        t_shift = float(self.config.get("t_shift", 0.1))
+        speed = float(self.config.get("speed", 1.0))
+        duration_value = float(self.config.get("duration", 0.0))
+        duration = duration_value if duration_value > 0 else None
+        layer_penalty_factor = float(self.config.get("layer_penalty_factor", 5.0))
+        position_temperature = float(self.config.get("position_temperature", 5.0))
+        class_temperature = float(self.config.get("class_temperature", 0.0))
+        denoise = bool(self.config.get("denoise", True))
+        preprocess_prompt = bool(self.config.get("preprocess_prompt", True))
+        postprocess_output = bool(self.config.get("postprocess_output", True))
+        audio_chunk_duration = float(self.config.get("audio_chunk_duration", 15.0))
+        audio_chunk_threshold = float(self.config.get("audio_chunk_threshold", 30.0))
+
+        cache_key = None
+        if enable_audio_cache:
+            cache_key = self.audio_cache.generate_cache_key(
+                "omnivoice",
+                text=stripped,
+                audio_component=audio_component,
+                prompt_text=prompt_text,
+                model_variant=model_variant,
+                language=language or "none",
+                instruct=instruct or "",
+                num_step=num_step,
+                guidance_scale=guidance_scale,
+                t_shift=t_shift,
+                speed=speed,
+                duration=duration if duration is not None else "auto",
+                layer_penalty_factor=layer_penalty_factor,
+                position_temperature=position_temperature,
+                class_temperature=class_temperature,
+                denoise=denoise,
+                preprocess_prompt=preprocess_prompt,
+                postprocess_output=postprocess_output,
+                audio_chunk_duration=audio_chunk_duration,
+                audio_chunk_threshold=audio_chunk_threshold,
+                seed=int(seed or 0),
+                device=device,
+                dtype=dtype,
+                character=character_name or "narrator",
+            )
+            cached_audio = self.audio_cache.get_cached_audio(cache_key)
+            if cached_audio:
+                print(f"💾 Using cached OmniVoice audio for '{character_name or 'narrator'}': '{stripped[:30]}...'")
+                return cached_audio[0]
+
+        voice_clone_prompt = None
+        if ref_audio is not None:
+            if not prompt_text:
+                raise ValueError(
+                    "OmniVoice voice cloning requires reference text. "
+                    "Direct audio-only narrator input is unsupported. "
+                    "Use Character Voices or a narrator voice file with matching reference text."
+                )
+            voice_clone_prompt = engine.create_voice_clone_prompt(
+                ref_audio=ref_audio,
+                ref_text=prompt_text,
+                preprocess_prompt=preprocess_prompt,
+            )
+
+        self._seed_everything(seed)
+
+        result = engine.generate(
+            text=stripped,
+            language=language,
+            voice_clone_prompt=voice_clone_prompt,
+            instruct=instruct,
+            duration=duration,
+            speed=speed,
+            num_step=num_step,
+            guidance_scale=guidance_scale,
+            t_shift=t_shift,
+            denoise=denoise,
+            preprocess_prompt=preprocess_prompt,
+            postprocess_output=postprocess_output,
+            layer_penalty_factor=layer_penalty_factor,
+            position_temperature=position_temperature,
+            class_temperature=class_temperature,
+            audio_chunk_duration=audio_chunk_duration,
+            audio_chunk_threshold=audio_chunk_threshold,
+        )
+
+        audio_tensor = result[0]
+        if not isinstance(audio_tensor, torch.Tensor):
+            audio_tensor = torch.tensor(audio_tensor, dtype=torch.float32)
+        audio_tensor = audio_tensor.detach().float().cpu()
+        if audio_tensor.dim() == 1:
+            audio_tensor = audio_tensor.unsqueeze(0)
+
+        if enable_audio_cache and cache_key:
+            duration_seconds = self.audio_cache._calculate_duration(audio_tensor, "omnivoice")
+            self.audio_cache.cache_audio(cache_key, audio_tensor, duration_seconds)
+
+        return audio_tensor

engines/adapters/qwen3_tts_adapter.py

@@ -19,7 +19,8 @@ if project_root not in sys.path:
 
 from engines.qwen3_tts.qwen3_tts import Qwen3TTSEngine
 from utils.text.pause_processor import PauseTagProcessor
-from utils.audio.cache import get_audio_cache
+from utils.audio.cache import get_audio_cache
+from utils.voice.reference import effective_voice_audio
 import folder_paths
 
 
@@ -105,9 +106,14 @@ class Qwen3TTSEngineAdapter:
         Returns:
             Model type string: "CustomVoice", "VoiceDesign", or "Base"
         """
-        # Priority 1: Voice Designer node
-        if context.get("node_type") == "voice_designer":
-            return "VoiceDesign"
+        # Explicit engine selection is authoritative for refactored workflows.
+        explicit_model_type = context.get("model_type")
+        if explicit_model_type in {"Base", "CustomVoice", "VoiceDesign"}:
+            return explicit_model_type
+
+        # Legacy voice designer context.
+        if context.get("node_type") == "voice_designer":
+            return "VoiceDesign"
 
         # Priority 2: Preset voice selected
         voice_preset = context.get("voice_preset")
@@ -117,15 +123,17 @@ class Qwen3TTSEngineAdapter:
         # Priority 3: Default to Base (zero-shot)
         return "Base"
 
-    def load_base_model(self,
-                       model_path: str,
-                       device: str = "auto",
-                       dtype: str = "auto",
-                       model_size: str = "1.7B",
-                       attn_implementation: str = "auto",
-                       context: Optional[Dict[str, Any]] = None,
-                       use_torch_compile: bool = False,
-                       use_cuda_graphs: bool = False,
+    def load_base_model(self,
+                       model_path: str,
+                       device: str = "auto",
+                       dtype: str = "auto",
+                       model_size: str = "1.7B",
+                       attn_implementation: str = "auto",
+                       runtime_mode: str = "main_environment",
+                       runtime_profile: Optional[str] = None,
+                       context: Optional[Dict[str, Any]] = None,
+                       use_torch_compile: bool = False,
+                       use_cuda_graphs: bool = False,
                        compile_mode: str = "reduce-overhead"):
         """
         Load Qwen3-TTS engine via unified interface with intelligent model selection.
@@ -153,8 +161,8 @@ class Qwen3TTSEngineAdapter:
             model_size = "1.7B"
             print("⚠️ VoiceDesign requires 1.7B model, auto-switching from 0.6B")
 
-        # Build model name
-        model_name = f"Qwen3-TTS-12Hz-{model_size}-{model_type}"
+        # Keep the canonical model name separate from a local: model path.
+        model_name = context.get("model_name") or f"Qwen3-TTS-12Hz-{model_size}-{model_type}"
 
         # Track current model type (unified interface handles unloading automatically)
         self.current_model_type = model_type
@@ -165,12 +173,14 @@ class Qwen3TTSEngineAdapter:
         config = ModelLoadConfig(
             engine_name="qwen3_tts",
             model_type="tts",
-            model_name=model_name,
-            model_path=model_path if model_path else model_name,
-            device=resolve_torch_device(device),
-            additional_params={
-                "dtype": dtype,
-                "attn_implementation": attn_implementation,
+            model_name=model_name,
+            model_path=model_path if model_path else model_name,
+            device=resolve_torch_device(device),
+            runtime_mode=runtime_mode,
+            runtime_profile=runtime_profile,
+            additional_params={
+                "dtype": dtype,
+                "attn_implementation": attn_implementation,
                 "use_torch_compile": use_torch_compile,
                 "use_cuda_graphs": use_cuda_graphs,
                 "compile_mode": compile_mode
@@ -524,20 +534,18 @@ class Qwen3TTSEngineAdapter:
         # Generate cache key using voice_ref dict (contains waveform + sample_rate)
         # This ensures different voices generate different cache keys
         from utils.audio.audio_hash import generate_stable_audio_component
-        if voice_ref and isinstance(voice_ref, dict):
-            # Check if voice_ref has audio tensor or file path
-            if "audio" in voice_ref:
-                # Unified Character Voices format: {"audio": {"waveform": ..., "sample_rate": ...}, "audio_path": ..., ...}
-                audio_dict = voice_ref.get("audio")
-                audio_component = generate_stable_audio_component(reference_audio=audio_dict)
-            elif ref_audio_original is not None and isinstance(ref_audio_original, str):
-                # File path format: {"audio_path": "/path/to/file.wav", "reference_text": "..."}
-                audio_component = generate_stable_audio_component(audio_file_path=ref_audio_original)
-            elif "waveform" in voice_ref:
-                # Direct tensor format: {"waveform": tensor, "sample_rate": 24000}
-                audio_component = generate_stable_audio_component(reference_audio=voice_ref)
-            else:
-                audio_component = "default_voice"
+        if voice_ref and isinstance(voice_ref, dict):
+            if isinstance(ref_audio_original, str):
+                audio_component = generate_stable_audio_component(audio_file_path=ref_audio_original)
+            elif isinstance(ref_audio_original, dict) and "waveform" in ref_audio_original:
+                audio_component = generate_stable_audio_component(reference_audio=ref_audio_original)
+            elif torch.is_tensor(ref_audio_original):
+                audio_component = generate_stable_audio_component(reference_audio={
+                    "waveform": ref_audio_original,
+                    "sample_rate": voice_ref.get("sample_rate", 24000),
+                })
+            else:
+                audio_component = "default_voice"
         elif ref_audio_original is not None and isinstance(ref_audio_original, str):
             # File path case (voice_ref is None but ref_audio_original extracted)
             audio_component = generate_stable_audio_component(audio_file_path=ref_audio_original)
@@ -683,10 +691,7 @@ class Qwen3TTSEngineAdapter:
             return None, None, False
 
         # Extract reference audio (multiple possible keys)
-        ref_audio_original = (voice_ref.get('prompt_audio_path') or
-                             voice_ref.get('audio_path') or
-                             voice_ref.get('audio') or
-                             voice_ref.get('waveform'))
+        ref_audio_original = effective_voice_audio(voice_ref)
 
         # Extract reference text (multiple possible keys)
         ref_text = (voice_ref.get('prompt_text') or

engines/adapters/step_audio_editx_adapter.py

@@ -93,7 +93,9 @@ class StepAudioEditXEngineAdapter:
                        model_path: str,
                        device: str = "auto",
                        torch_dtype: str = "auto",
-                       quantization: Optional[str] = None):
+                       quantization: Optional[str] = None,
+                       runtime_mode: str = "shared_runtime",
+                       runtime_profile: Optional[str] = None):
         """
         Load Step Audio EditX engine via unified interface (with caching).
 
@@ -113,6 +115,8 @@ class StepAudioEditXEngineAdapter:
             model_name="Step-Audio-EditX",
             model_path=model_path,  # Downloader will resolve this if it's "local:xxx"
             device=resolve_torch_device(device),
+            runtime_mode=runtime_mode,
+            runtime_profile=runtime_profile,
             additional_params={
                 "torch_dtype": torch_dtype,
                 "quantization": quantization
@@ -193,7 +197,7 @@ class StepAudioEditXEngineAdapter:
             prompt_text=prompt_text,
             temperature=params.get('temperature', 0.7),
             do_sample=params.get('do_sample', True),
-            max_new_tokens=params.get('max_new_tokens', 8192),
+            max_new_tokens=params.get('max_new_tokens', 1024),
             seed=params.get('seed', 0),
             model_path=params.get('model_path', 'Step-Audio-EditX'),
             device=params.get('device', 'auto'),
@@ -210,7 +214,7 @@ class StepAudioEditXEngineAdapter:
             return cached_audio[0]
 
         # Create ComfyUI progress bar for generation tracking with time prediction
-        max_new_tokens = params.get('max_new_tokens', 8192)
+        max_new_tokens = params.get('max_new_tokens', 1024)
 
         # Estimate actual tokens based on text length (same heuristic as Qwen3-TTS)
         # Rough heuristic: ~0.7 tokens per character for TTS (conservative estimate)

engines/adapters/vibevoice_adapter.py

@@ -16,6 +16,8 @@ if parent_dir not in sys.path:
     sys.path.insert(0, parent_dir)
 
 from utils.models.language_mapper import get_model_for_language
+from utils.models.engine_registry import get_default_runtime_profile
+from utils.models.factory_config import normalize_runtime_mode, runtime_uses_isolation
 from utils.text.pause_processor import PauseTagProcessor
 from utils.text.character_parser import character_parser
 from engines.vibevoice_engine.vibevoice_downloader import (
@@ -23,6 +25,7 @@ from engines.vibevoice_engine.vibevoice_downloader import (
     is_kugelaudio_variant_name,
 )
 from utils.models.manager import model_manager
+from utils.voice.character_logging import resolved_character_label
 
 
 class VibeVoiceEngineAdapter:
@@ -51,11 +54,15 @@ class VibeVoiceEngineAdapter:
             self._device = engine_config.get('device', 'auto')
             self._attention_mode = engine_config.get('attention_mode', 'auto')
             self._quantize_4bit = engine_config.get('quantize_llm_4bit', False)
+            self._runtime_mode = normalize_runtime_mode(engine_config.get('runtime_mode'))
+            self._runtime_profile = engine_config.get('runtime_profile')
         else:
             self._model_name = 'vibevoice-1.5B'
             self._device = 'auto'
             self._attention_mode = 'auto'
             self._quantize_4bit = False
+            self._runtime_mode = 'main_environment'
+            self._runtime_profile = None
 
     @property
     def vibevoice_engine(self):
@@ -68,7 +75,9 @@ class VibeVoiceEngineAdapter:
             model_name=self._model_name,
             device=self._device,
             attention_mode=self._attention_mode,
-            quantize_llm_4bit=self._quantize_4bit
+            quantize_llm_4bit=self._quantize_4bit,
+            runtime_mode=self._runtime_mode,
+            runtime_profile=self._runtime_profile or get_default_runtime_profile("vibevoice"),
         )
 
     @property
@@ -458,7 +467,8 @@ class VibeVoiceEngineAdapter:
         
         # Generate each character group using VibeVoice format
         for group_idx, (character, text_list) in enumerate(character_groups):
-            print(f"🎤 Group {group_idx + 1}: Character '{character}' with {len(text_list)} segments")
+            display_name = resolved_character_label(character, voice_mapping.get(character))
+            print(f"🎤 Group {group_idx + 1}: Character '{display_name}' with {len(text_list)} segments")
             
             # Format as Speaker 1 entries (VibeVoice style) and combine
             formatted_lines = []
@@ -518,15 +528,11 @@ class VibeVoiceEngineAdapter:
         Returns:
             Combined audio dict
         """
-        # Get speaker voice inputs from engine config for priority system
-        # For manual Speaker format, the main narrator voice comes from the TTS Text node
-        # We need to get it from a different source since voice_mapping may not have 'narrator' key
-        main_narrator_voice = voice_mapping.get('narrator')  # Try narrator first
-        if main_narrator_voice is None:
-            # If no 'narrator' key, get it from any available voice (fallback for manual Speaker format)
-            available_voices = [v for v in voice_mapping.values() if v is not None]
-            main_narrator_voice = available_voices[0] if available_voices else None
-            # print(f"🐛 Debug: No 'narrator' key, using fallback voice: {'✅ found' if main_narrator_voice else '❌ none available'}")
+        # Get speaker voice inputs from engine config for priority system.
+        # Speaker 1 must only come from an explicit narrator/speaker-1 input.
+        # Do not synthesize Speaker 1 from discovered character voices, or aliases
+        # will look like they were overridden by a connection that does not exist.
+        main_narrator_voice = voice_mapping.get('narrator')
         speaker_inputs = {
             1: main_narrator_voice,  # Speaker 1 uses main narrator from TTS Text
             2: params.get('speaker2_voice'),
@@ -536,19 +542,24 @@ class VibeVoiceEngineAdapter:
         
         # print(f"🐛 Debug: speaker_inputs[1] (main narrator): {'✅ has voice' if speaker_inputs[1] else '❌ no voice'}")
         
-        # Build speaker mapping and format text
+        # Build speaker mapping and format text.
+        # For SRT/global processing we keep speaker numbering local to the current
+        # generation call (Speaker 1..N in order of first appearance), but preserve
+        # the character's global slot when choosing connected speaker override inputs.
         character_map = {}
-        # Pre-fill speaker_voices with all 4 speaker slots (some may be None)
-        speaker_voices = [
-            speaker_inputs.get(1),  # Speaker 1
-            speaker_inputs.get(2),  # Speaker 2
-            speaker_inputs.get(3),  # Speaker 3
-            speaker_inputs.get(4)   # Speaker 4
-        ]
+        character_global_slots = {}
+        speaker_voices = []
         formatted_lines = []
+        segment_characters = [char for char, _ in segments]
 
-        print(f"🎭 Native multi-speaker: Processing {len(segments)} segments with characters: {[char for char, _ in segments]}")
+        print(f"🎭 Native multi-speaker: Processing {len(segments)} segments with characters: {segment_characters}")
         print(f"🎤 Speaker inputs connected: {[f'Speaker {k}' for k, v in speaker_inputs.items() if v is not None]}")
+        narrator_exists_globally = bool(global_char_to_speaker and "narrator" in global_char_to_speaker)
+        if speaker_inputs.get(1) is not None and "narrator" not in segment_characters:
+            if narrator_exists_globally:
+                print("ℹ️ Narrator input is connected, but this subtitle has no narrator turn; it will not override tagged characters")
+            else:
+                print("ℹ️ No narrator turns exist in this SRT; Speaker 1 maps to the first named character by first appearance")
         
         for character, text in segments:
             # Check if this is already a manual "Speaker N:" format
@@ -559,50 +570,84 @@ class VibeVoiceEngineAdapter:
                 speaker_idx = manual_speaker - 1  # Convert to 0-based
                 if speaker_idx >= 4:
                     speaker_idx = 3
-                    
-                # Voice already in pre-filled speaker_voices array
-                voice = speaker_voices[speaker_idx]
+                voice = speaker_inputs.get(manual_speaker)
                 if manual_speaker == 1:
                     print(f"🎤 Manual format 'Speaker {manual_speaker}' -> using {'✅ main narrator (Tony)' if voice else '❌ no narrator, using default'}")
                 else:
                     print(f"🎤 Manual format 'Speaker {manual_speaker}' -> using {'✅ connected input' if voice else '❌ no input, using default'}")
 
+                while len(speaker_voices) <= speaker_idx:
+                    speaker_voices.append(None)
+                speaker_voices[speaker_idx] = voice
+
                 formatted_lines.append(f"Speaker {manual_speaker}: {text.strip()}")
                 
             else:
-                # Character tag format - use global mapping if provided (for SRT consistency)
+                # Character tag format
                 if character not in character_map:
                     # Special handling for numeric characters: [1] [2] [3] [4] -> map directly to Speaker N
                     if character.isdigit() and 1 <= int(character) <= 4:
-                        speaker_idx = int(character) - 1  # Convert [1] to Speaker 1 (0-based index)
+                        global_speaker_num = int(character)
+                        speaker_idx = len(character_map)
                         character_map[character] = speaker_idx
+                        character_global_slots[character] = global_speaker_num
                         print(f"🔢 Numeric character '[{character}]' -> Speaker {int(character)} (direct mapping)")
                     elif global_char_to_speaker and character in global_char_to_speaker:
-                        # Use global mapping for consistent SRT processing
-                        speaker_idx = global_char_to_speaker[character] - 1  # Convert to 0-based
+                        global_speaker_num = global_char_to_speaker[character]
+                        speaker_idx = len(character_map)
                         character_map[character] = speaker_idx
+                        character_global_slots[character] = global_speaker_num
                     else:
                         # Fallback to sequential assignment
                         speaker_idx = len(character_map)
                         if speaker_idx >= 4:
                             print(f"⚠️ VibeVoice: Limiting to 4 speakers, '{character}' will use Speaker 4")
                             speaker_idx = 3  # Use 0-based internally, will convert to 1-based for format
+                            global_speaker_num = 4
                         else:
                             character_map[character] = speaker_idx
+                            global_speaker_num = speaker_idx + 1
+                            character_global_slots[character] = global_speaker_num
                         
-                    # Priority system: speaker inputs override character aliases
+                    # Priority system:
+                    # - Numeric [1]-[4] always map directly to speaker inputs 1-4.
+                    # - If any narrator turn exists in the SRT, Speaker 1 is reserved for narrator.
+                    # - Otherwise, Speaker 1..4 map to named characters by first global appearance order.
                     speaker_num = speaker_idx + 1
-                    connected_voice = speaker_inputs.get(speaker_num)
+                    global_speaker_num = character_global_slots.get(character, speaker_num)
+                    connected_voice = None
+                    is_numeric_direct = character.isdigit() and 1 <= int(character) <= 4
+                    if is_numeric_direct:
+                        connected_voice = speaker_inputs.get(global_speaker_num)
+                    elif character == "narrator":
+                        connected_voice = speaker_inputs.get(1)
+                    elif narrator_exists_globally:
+                        if global_speaker_num >= 2:
+                            connected_voice = speaker_inputs.get(global_speaker_num)
+                    else:
+                        connected_voice = speaker_inputs.get(global_speaker_num)
                     character_voice = voice_mapping.get(character)
-                    
+
                     if connected_voice is not None and character_voice is not None:
-                        print(f"⚠️ Priority: Speaker {speaker_num} input overrides ['{character}'] alias - using connected voice")
+                        if global_speaker_num != speaker_num:
+                            print(
+                                f"⚠️ Priority: Speaker {global_speaker_num} input overrides ['{character}'] alias "
+                                f"- using connected voice as local Speaker {speaker_num}"
+                            )
+                        else:
+                            print(f"⚠️ Priority: Speaker {speaker_num} input overrides ['{character}'] alias - using connected voice")
                         voice = connected_voice
                     elif connected_voice is not None:
-                        print(f"🎤 Speaker {speaker_num}: Using connected voice input")
+                        if global_speaker_num != speaker_num:
+                            print(f"🎤 Speaker {global_speaker_num}: Using connected voice input as local Speaker {speaker_num}")
+                        else:
+                            print(f"🎤 Speaker {speaker_num}: Using connected voice input")
                         voice = connected_voice
                     else:
-                        print(f"🎭 Character '{character}' -> Speaker {speaker_num}, using character voice")
+                        if global_speaker_num != speaker_num:
+                            print(f"🎭 Character '{character}' -> global Speaker {global_speaker_num}, local Speaker {speaker_num}, using alias/character voice")
+                        else:
+                            print(f"🎭 Character '{character}' -> Speaker {speaker_num}, using alias/character voice")
                         voice = character_voice
                     
                     # Ensure we have enough speaker_voices slots
@@ -621,6 +666,18 @@ class VibeVoiceEngineAdapter:
         print(formatted_text)
         print("="*60)
         print(f"🎤 Using {len(speaker_voices)} voice samples for generation")
+        for idx, voice in enumerate(speaker_voices, start=1):
+            if voice is None:
+                print(f"   Speaker {idx}: default / no reference")
+            elif isinstance(voice, dict) and voice.get("audio_path"):
+                print(f"   Speaker {idx}: file reference -> {voice['audio_path']}")
+            elif isinstance(voice, dict) and "waveform" in voice:
+                waveform = voice["waveform"]
+                sample_rate = voice.get("sample_rate", "unknown")
+                shape = tuple(waveform.shape) if hasattr(waveform, "shape") else "unknown"
+                print(f"   Speaker {idx}: waveform reference -> shape {shape}, sr {sample_rate}")
+            else:
+                print(f"   Speaker {idx}: unexpected reference type {type(voice).__name__}")
         
         # Validate and normalize voice references
         normalized_voices = []
@@ -847,7 +904,14 @@ class VibeVoiceEngineAdapter:
         return None
     
     def cleanup(self):
-        """Clean up resources - models are now managed by ModelManager"""
+        """Clean up resources and unload isolated runtime workers when requested."""
         self._character_speaker_map.clear()
         self._speaker_voices.clear()
+
+        if runtime_uses_isolation(self._runtime_mode):
+            try:
+                from utils.models.unified_model_interface import unified_model_interface
+                unified_model_interface.clear_engine_models("vibevoice")
+            except Exception as e:
+                print(f"⚠️ Failed to unload isolated VibeVoice runtime during cleanup: {e}")
         # Don't clean up vibevoice_engine - it just holds references to cached models

engines/chatterbox/audio_timing.py

@@ -18,12 +18,28 @@ import shutil
 import soundfile as sf
 
 
-class AudioTimingError(Exception):
-    """Exception raised when audio timing operations fail"""
-    pass
-
-
-class AudioTimingUtils:
+class AudioTimingError(Exception):
+    """Exception raised when audio timing operations fail"""
+    pass
+
+
+def _stack_stretched_channels(
+    channels: List[torch.Tensor],
+    device: torch.device,
+) -> torch.Tensor:
+    """Stack independently stretched channels after reconciling tiny length drift."""
+    if not channels:
+        raise AudioTimingError("No audio channels were processed successfully")
+    common_length = min(channel.size(-1) for channel in channels)
+    if common_length <= 0:
+        raise AudioTimingError("Time stretching produced an empty audio channel")
+    return torch.stack(
+        [channel[..., :common_length] for channel in channels],
+        dim=0,
+    ).to(device)
+
+
+class AudioTimingUtils:
     """
     Utilities for audio timing manipulation and synchronization
     """
@@ -211,7 +227,7 @@ class PhaseVocoderTimeStretcher:
                 stretched_channels.append(torch.from_numpy(stretched))
         
         # Combine channels
-        result = torch.stack(stretched_channels, dim=0).to(audio.device)
+        result = _stack_stretched_channels(stretched_channels, audio.device)
         
         # Restore original shape if input was 1D
         if len(original_shape) == 1:
@@ -366,7 +382,7 @@ class FFmpegTimeStretcher:
             if not stretched:
                 raise AudioTimingError("No audio was processed successfully")
                 
-            result = torch.stack(stretched, dim=0).to(audio.device)
+            result = _stack_stretched_channels(stretched, audio.device)
             return result.squeeze(0) if len(original_shape) == 1 else result
             
         except Exception as e:
@@ -425,7 +441,7 @@ class FFmpegTimeStretcher:
             
         try:
             # Stack channels and restore shape
-            result = torch.stack(stretched, dim=0).to(audio.device)
+            result = _stack_stretched_channels(stretched, audio.device)
             print(f"Successfully processed all channels")
             return result.squeeze(0) if len(original_shape) == 1 else result
             
@@ -707,4 +723,4 @@ def calculate_timing_adjustments(natural_durations: List[float],
         
         adjustments.append(adjustment)
     
-    return adjustments
+    return adjustments

engines/chatterbox/batch_processor.py

@@ -7,7 +7,8 @@ Works with character_grouper.py to process multiple segments simultaneously.
 
 from typing import Dict, List, Any
 import torch
-from .character_grouper import CharacterGroup
+from .character_grouper import CharacterGroup
+from utils.voice.character_logging import resolved_character_label
 
 
 class BatchProcessor:
@@ -49,7 +50,8 @@ class BatchProcessor:
         character = character_group.character
         segments = character_group.segments
         
-        print(f"🚀 BATCH PROCESSING: {character} - {len(segments)} segments in {language}")
+        display_name = resolved_character_label(character, voice_refs.get(character))
+        print(f"🚀 BATCH PROCESSING: {display_name} - {len(segments)} segments in {language}")
         
         # Collect all texts for batching
         batch_texts = []
@@ -75,7 +77,7 @@ class BatchProcessor:
         char_audio_prompt = voice_refs[character]
         
         # THE ACTUAL BATCH PROCESSING - this is the key improvement
-        print(f"⚡ Batch generating {len(batch_texts)} chunks for {character}")
+        print(f"⚡ Batch generating {len(batch_texts)} chunks for {display_name}")
         print(f"🔧 DEBUG: batch_size from inputs = {inputs.get('batch_size', 4)}")
         try:
             batch_audio = self.tts_model.generate_batch(
@@ -137,15 +139,15 @@ class BatchProcessor:
         character = character_group.character
         segments = character_group.segments
         
-        print(f"→ SEQUENTIAL: {character} - {len(segments)} segments in {language}")
+        char_audio_prompt = voice_refs[character]
+        display_name = resolved_character_label(character, char_audio_prompt)
+        print(f"→ SEQUENTIAL: {display_name} - {len(segments)} segments in {language}")
         
         results = {}
-        char_audio_prompt = voice_refs[character]
-        
-        for segment in segments:
+        for segment in segments:
             segment_display_idx = segment.original_idx + 1  # 1-based for display
             
-            print(f"🎤 Generating segment {segment_display_idx}/{total_segments} for '{character}' (lang: {language})")
+            print(f"🎤 Generating segment {segment_display_idx}/{total_segments} for '{display_name}' (lang: {language})")
             
             # Apply chunking
             if inputs["enable_chunking"] and len(segment.segment_text) > inputs["max_chars_per_chunk"]:
@@ -183,4 +185,4 @@ class BatchProcessor:
         """Apply crash protection padding to short texts."""
         if len(text.strip()) < min_length:
             return padding_template.format(seg=text)
-        return text
+        return text

engines/chatterbox_official_23lang/batch_processor.py

@@ -7,7 +7,8 @@ Works with character_grouper.py to process multiple segments simultaneously.
 
 from typing import Dict, List, Any
 import torch
-from .character_grouper import CharacterGroup
+from .character_grouper import CharacterGroup
+from utils.voice.character_logging import resolved_character_label
 
 
 class BatchProcessor:
@@ -49,7 +50,8 @@ class BatchProcessor:
         character = character_group.character
         segments = character_group.segments
         
-        print(f"🚀 BATCH PROCESSING: {character} - {len(segments)} segments in {language}")
+        display_name = resolved_character_label(character, voice_refs.get(character))
+        print(f"🚀 BATCH PROCESSING: {display_name} - {len(segments)} segments in {language}")
         
         # Collect all texts for batching
         batch_texts = []
@@ -76,7 +78,7 @@ class BatchProcessor:
         char_audio_prompt = voice_refs[character]
         
         # THE ACTUAL BATCH PROCESSING - this is the key improvement
-        print(f"⚡ Batch generating {len(batch_texts)} chunks for {character}")
+        print(f"⚡ Batch generating {len(batch_texts)} chunks for {display_name}")
         print(f"🔧 DEBUG: batch_size from inputs = {inputs.get('batch_size', 4)}")
         try:
             batch_audio = self.tts_model.generate_batch(
@@ -138,15 +140,15 @@ class BatchProcessor:
         character = character_group.character
         segments = character_group.segments
         
-        print(f"→ SEQUENTIAL: {character} - {len(segments)} segments in {language}")
+        char_audio_prompt = voice_refs[character]
+        display_name = resolved_character_label(character, char_audio_prompt)
+        print(f"→ SEQUENTIAL: {display_name} - {len(segments)} segments in {language}")
         
         results = {}
-        char_audio_prompt = voice_refs[character]
-        
-        for segment in segments:
+        for segment in segments:
             segment_display_idx = segment.original_idx + 1  # 1-based for display
             
-            print(f"🎤 Generating segment {segment_display_idx}/{total_segments} for '{character}' (lang: {language})")
+            print(f"🎤 Generating segment {segment_display_idx}/{total_segments} for '{display_name}' (lang: {language})")
             
             # Apply chunking
             if inputs["enable_chunking"] and len(segment.segment_text) > inputs["max_chars_per_chunk"]:
@@ -184,4 +186,4 @@ class BatchProcessor:
         """Apply crash protection padding to short texts."""
         if len(text.strip()) < min_length:
             return padding_template.format(seg=text)
-        return text
+        return text

engines/chatterbox_official_23lang/language_models.py

@@ -43,19 +43,30 @@ OFFICIAL_23LANG_MODELS = {
         "required_files": {
             "v1": [
                 "t3_23lang.safetensors",  # Multilingual T3 model v1
-                "s3gen.pt",               # S3Gen model (same as English)
-                "ve.pt",                  # Voice encoder (same as English)
-                "mtl_tokenizer.json",     # Multilingual tokenizer
-                "conds.pt"                # Conditioning (optional)
+                "s3gen.pt",               # S3Gen model (same as English)
+                "ve.pt",                  # Voice encoder (same as English)
+                "mtl_tokenizer.json",     # Multilingual tokenizer
+                "Cangjie5_TC.json",       # Chinese Cangjie mapping
+                "conds.pt"                # Conditioning (optional)
             ],
-            "v2": [
-                "t3_mtl23ls_v2.safetensors",      # Multilingual T3 model v2 with enhanced tokenization
-                "s3gen.pt",                        # S3Gen model (same as English)
-                "ve.pt",                           # Voice encoder (same as English)
-                "grapheme_mtl_merged_expanded_v1.json",  # Enhanced grapheme/phoneme mappings with special tokens
-                "mtl_tokenizer.json",              # Multilingual tokenizer (may be updated for v2)
-                "conds.pt"                         # Conditioning (optional)
-            ]
+            "v2": [
+                "t3_mtl23ls_v2.safetensors",      # Multilingual T3 model v2 with enhanced tokenization
+                "s3gen.pt",                        # S3Gen model (same as English)
+                "ve.pt",                           # Voice encoder (same as English)
+                "grapheme_mtl_merged_expanded_v1.json",  # Enhanced grapheme/phoneme mappings with special tokens
+                "mtl_tokenizer.json",              # Multilingual tokenizer (may be updated for v2)
+                "Cangjie5_TC.json",                 # Chinese Cangjie mapping
+                "conds.pt"                         # Conditioning (optional)
+            ],
+            "v3": [
+                "t3_mtl23ls_v3.safetensors",      # Latest official multilingual T3 model
+                "s3gen.pt",                        # Official V3 API continues to use shared S3Gen
+                "ve.pt",                           # Shared voice encoder
+                "grapheme_mtl_merged_expanded_v1.json",
+                "mtl_tokenizer.json",
+                "Cangjie5_TC.json",
+                "conds.pt"
+            ]
         },
         "multilingual": True
     },

engines/chatterbox_official_23lang/models/t3/t3.py

@@ -235,6 +235,9 @@ class T3(nn.Module):
         length_penalty=1.0,
         repetition_penalty=1.2,
         cfg_weight=0.5,
+        # TTS Audio Suite patch: V3 follows upstream by disabling the legacy
+        # multilingual alignment analyzer while V1/V2 retain existing behavior.
+        use_alignment_analyzer=True,
     ):
         """
         Args:
@@ -267,7 +270,7 @@ class T3(nn.Module):
         if not self.compiled:
             # Default to None for English models, only create for multilingual
             alignment_stream_analyzer = None
-            if self.hp.is_multilingual:
+            if self.hp.is_multilingual and use_alignment_analyzer:
                 alignment_stream_analyzer = AlignmentStreamAnalyzer(
                     self.tfmr,
                     None,
@@ -331,7 +334,7 @@ class T3(nn.Module):
             inputs_embeds=inputs_embeds,
             past_key_values=None,
             use_cache=True,
-            output_attentions=True,
+            output_attentions=use_alignment_analyzer,
             output_hidden_states=True,
             return_dict=True,
         )

engines/chatterbox_official_23lang/models/tokenizers/tokenizer.py

@@ -6,7 +6,6 @@ import torch
 from pathlib import Path
 from unicodedata import category
 from tokenizers import Tokenizer
-from huggingface_hub import hf_hub_download
 from utils.text.russian_stress_support import get_russian_text_stresser
 
 
@@ -56,9 +55,6 @@ class EnTokenizer:
         return txt
 
 
-# Model repository
-REPO_ID = "ResembleAI/chatterbox"
-
 # Global instances for optional dependencies
 _kakasi = None
 _dicta = None
@@ -167,13 +163,13 @@ class ChineseCangjieConverter:
         self._init_segmenter()
     
     def _load_cangjie_mapping(self, model_dir=None):
-        """Load Cangjie mapping from HuggingFace model repository."""        
+        """Load the Cangjie mapping from the organized local model folder."""
         try:
-            cangjie_file = hf_hub_download(
-                repo_id=REPO_ID,
-                filename="Cangjie5_TC.json",
-                cache_dir=model_dir
-            )
+            # TTS Audio Suite patch: this asset is downloaded by the unified
+            # downloader; tokenization must never create a hidden HF cache.
+            cangjie_file = Path(model_dir) / "Cangjie5_TC.json"
+            if not cangjie_file.is_file():
+                raise FileNotFoundError(f"Missing local Cangjie mapping: {cangjie_file}")
             
             with open(cangjie_file, "r", encoding="utf-8") as fp:
                 data = json.load(fp)

engines/chatterbox_official_23lang/tts.py

@@ -29,7 +29,7 @@ except ImportError:
     PERTH_AVAILABLE = False
 
 from .models.t3 import T3
-from .models.s3tokenizer import S3_SR, drop_invalid_tokens
+from .models.s3tokenizer import S3_SR, S3_TOKEN_RATE, drop_invalid_tokens
 from .models.s3gen import S3GEN_SR, S3Gen
 from .models.tokenizers import EnTokenizer, MTLTokenizer
 from .models.voice_encoder import VoiceEncoder
@@ -219,7 +219,8 @@ class ChatterboxOfficial23LangTTS:
         """
         Load ChatterBox Official 23-Lang multilingual model from local directory.
         Expected files:
-        - t3_23lang.safetensors (multilingual T3 model v1) OR t3_mtl23ls_v2.safetensors (v2)
+        - t3_23lang.safetensors (v1), t3_mtl23ls_v2.safetensors (v2),
+          or t3_mtl23ls_v3.safetensors (v3)
         - s3gen.pt (S3Gen model)
         - ve.pt (Voice encoder)
         - mtl_tokenizer.json (multilingual tokenizer)
@@ -281,8 +282,8 @@ class ChatterboxOfficial23LangTTS:
             print("📦 Loading multilingual tokenizer...")
             tokenizer_path = None
 
-            if version_for_files == "v2":
-                # Try v2 enhanced tokenizer first
+            if version_for_files in ("v2", "v3"):
+                # V2 and V3 use the expanded multilingual tokenizer.
                 candidate_path = ckpt_dir / "grapheme_mtl_merged_expanded_v1.json"
                 if candidate_path.exists():
                     tokenizer_path = candidate_path
@@ -327,23 +328,26 @@ class ChatterboxOfficial23LangTTS:
 
             # Support multiple T3 filename patterns:
             # - Official v1: t3_23lang.safetensors
-            # - Official v2: t3_mtl23ls_v2.safetensors
+            # - Official v2/v3: t3_mtl23ls_v2.safetensors / t3_mtl23ls_v3.safetensors
             # - Vietnamese Viterbox: t3_ml24ls_v2.safetensors
             # - Egyptian Arabic: t3_mtl23ls_v2.safetensors
             # - Future variants: any t3_*.safetensors
             t3_path = None
-            if version_for_files == "v2":
-                # Try specific v2 patterns first
-                for pattern in ["t3_mtl23ls_v2.safetensors", "t3_ml24ls_v2.safetensors"]:
+            if version_for_files in ("v2", "v3"):
+                patterns = (
+                    ["t3_mtl23ls_v3.safetensors"]
+                    if version_for_files == "v3"
+                    else ["t3_mtl23ls_v2.safetensors", "t3_ml24ls_v2.safetensors"]
+                )
+                for pattern in patterns:
                     candidate = ckpt_dir / pattern
                     if candidate.exists():
                         t3_path = candidate
                         break
 
-                # Fallback: find any t3_*_v2.safetensors file
-                if not t3_path:
-                    import glob
-                    matches = list(ckpt_dir.glob("t3_*_v2.safetensors"))
+                # Fallback stays version-specific so V2 and V3 cannot be mixed.
+                if not t3_path:
+                    matches = list(ckpt_dir.glob(f"t3_*_{version_for_files}.safetensors"))
                     if matches:
                         t3_path = matches[0]
             else:
@@ -505,7 +509,7 @@ class ChatterboxOfficial23LangTTS:
         Args:
             device: Device to load model on
             model_name: Model to load (defaults to "ChatterBox Official 23-Lang")
-            model_version: Model version - "v1" or "v2" (defaults to "v2")
+            model_version: Model version - "v1", "v2", or "v3"
         """
         # Get model configuration
         model_config = get_model_config(model_name)
@@ -689,9 +693,10 @@ class ChatterboxOfficial23LangTTS:
                 max_new_tokens=1000,  # TODO: use the value in config
                 temperature=temperature,
                 cfg_weight=cfg_weight,
-                repetition_penalty=repetition_penalty,
-                min_p=min_p,
-                top_p=top_p,
+                repetition_penalty=repetition_penalty,
+                min_p=min_p,
+                top_p=top_p,
+                use_alignment_analyzer=self.model_version != "v3",
             )
             # Extract only the conditional batch.
             speech_tokens = speech_tokens[0]
@@ -700,12 +705,20 @@ class ChatterboxOfficial23LangTTS:
             speech_tokens = drop_invalid_tokens(speech_tokens)
             speech_tokens = speech_tokens.to(self.device)
 
-            wav, _ = self.s3gen.inference(
-                speech_tokens=speech_tokens,
-                ref_dict=self.conds.gen,
-            )
-            wav = wav.squeeze(0).detach().cpu().numpy()
-            if self.enable_watermarking:
+            wav, _ = self.s3gen.inference(
+                speech_tokens=speech_tokens,
+                ref_dict=self.conds.gen,
+            )
+            wav = wav.squeeze(0).detach().cpu().numpy()
+
+            if self.model_version == "v3":
+                # TTS Audio Suite patch: match official V3 by dropping the
+                # final degraded pre-EOS speech-token artifact.
+                token_count = int(speech_tokens.shape[-1])
+                clean_token_count = max(1, token_count - 1)
+                wav = wav[: clean_token_count * (S3GEN_SR // S3_TOKEN_RATE)]
+
+            if self.enable_watermarking:
                 self._init_watermarker_if_needed()
                 if self.watermarker is not None:
                     watermarked_wav = self.watermarker.apply_watermark(wav, sample_rate=self.sr)

engines/cosyvoice/impl/cosyvoice/cli/cosyvoice.py

@@ -270,7 +270,6 @@ class CosyVoice3(CosyVoice2):
     # PATCH: Added llm_filename parameter to support model variants (llm.pt vs llm.rl.pt)
     def __init__(self, model_dir, load_trt=False, load_vllm=False, fp16=False, trt_concurrent=1, llm_filename='llm.pt'):
         self.model_dir = model_dir
-        self.fp16 = fp16
         if not os.path.exists(model_dir):
             model_dir = snapshot_download(model_dir)
         hyper_yaml_path = '{}/cosyvoice3.yaml'.format(model_dir)
@@ -289,6 +288,15 @@ class CosyVoice3(CosyVoice2):
         if torch.cuda.is_available() is False and (load_trt is True or fp16 is True):
             load_trt, fp16 = False, False
             logging.warning('no cuda device, set load_trt/fp16 to False')
+        # TTS Audio Suite patch: ROCm cannot use CosyVoice3's FP16 flow/vocoder
+        # path reliably; CosyVoice3Model applies BF16 only to the bundled Qwen LLM.
+        elif torch.version.hip:
+            if fp16:
+                logging.warning('ROCm detected: CosyVoice3 disables FP16 for its flow/vocoder and uses BF16 only for the Qwen LLM.')
+            if load_trt:
+                logging.warning('ROCm detected: TensorRT is unavailable, disabling CosyVoice3 TensorRT loading.')
+            load_trt, fp16 = False, False
+        self.fp16 = fp16
         self.model = CosyVoice3Model(configs['llm'], configs['flow'], configs['hift'], fp16)
         # PATCH: Use llm_filename parameter instead of hardcoded 'llm.pt'
         self.model.load('{}/{}'.format(model_dir, llm_filename),

engines/cosyvoice/impl/cosyvoice/cli/model.py

@@ -38,6 +38,7 @@ class CosyVoiceModel:
         self.flow = flow
         self.hift = hift
         self.fp16 = fp16
+        self._rocm_bf16_llm_autocast = False
         self.token_min_hop_len = 2 * self.flow.input_frame_rate
         self.token_max_hop_len = 4 * self.flow.input_frame_rate
         self.token_overlap_len = 20
@@ -57,6 +58,7 @@ class CosyVoiceModel:
         # dict used to store session related variable
         self.tts_speech_token_dict = {}
         self.llm_end_dict = {}
+        self.llm_error_dict = {}
         self.mel_overlap_dict = {}
         self.flow_cache_dict = {}
         self.hift_cache_dict = {}
@@ -111,29 +113,61 @@ class CosyVoiceModel:
         input_names = ["x", "mask", "mu", "t", "spks", "cond"]
         return {'min_shape': min_shape, 'opt_shape': opt_shape, 'max_shape': max_shape, 'input_names': input_names}
 
+    # TTS Audio Suite patch: ROCm needs BF16 for the bundled Qwen LLM while
+    # CosyVoice's flow model and vocoder remain in FP32 to prevent silent audio.
+    def _llm_autocast_context(self):
+        if hasattr(self.llm, 'vllm'):
+            return nullcontext()
+        if self._rocm_bf16_llm_autocast:
+            return torch.amp.autocast('cuda', dtype=torch.bfloat16)
+        if self.fp16:
+            return torch.amp.autocast('cuda', dtype=torch.float16)
+        return nullcontext()
+
+    # TTS Audio Suite patch: forward bundled LLM worker failures to the caller
+    # instead of trying to decode an empty token sequence.
+    def _raise_if_llm_failed(self, uuid):
+        error = self.llm_error_dict.pop(uuid, None)
+        if error is None:
+            return
+        with self.lock:
+            self.tts_speech_token_dict.pop(uuid, None)
+            self.llm_end_dict.pop(uuid, None)
+            self.hift_cache_dict.pop(uuid, None)
+            if hasattr(self, 'mel_overlap_dict'):
+                self.mel_overlap_dict.pop(uuid, None)
+            if hasattr(self, 'flow_cache_dict'):
+                self.flow_cache_dict.pop(uuid, None)
+        raise RuntimeError(f'CosyVoice LLM inference failed: {error}') from error
+
     def llm_job(self, text, prompt_text, llm_prompt_speech_token, llm_embedding, uuid, progress_callback=None):
-        with self.llm_context, torch.cuda.amp.autocast(self.fp16 is True and hasattr(self.llm, 'vllm') is False):
-            if isinstance(text, Generator):
-                assert (self.__class__.__name__ != 'CosyVoiceModel') and not hasattr(self.llm, 'vllm'), 'streaming input text is only implemented for CosyVoice2/3 and do not support vllm!'
-                for i in self.llm.inference_bistream(text=text,
-                                                     prompt_text=prompt_text.to(self.device),
-                                                     prompt_text_len=torch.tensor([prompt_text.shape[1]], dtype=torch.int32).to(self.device),
-                                                     prompt_speech_token=llm_prompt_speech_token.to(self.device),
-                                                     prompt_speech_token_len=torch.tensor([llm_prompt_speech_token.shape[1]], dtype=torch.int32).to(self.device),
-                                                     embedding=llm_embedding.to(self.device)):
-                    self.tts_speech_token_dict[uuid].append(i)
-            else:
-                for i in self.llm.inference(text=text.to(self.device),
-                                            text_len=torch.tensor([text.shape[1]], dtype=torch.int32).to(self.device),
-                                            prompt_text=prompt_text.to(self.device),
-                                            prompt_text_len=torch.tensor([prompt_text.shape[1]], dtype=torch.int32).to(self.device),
-                                            prompt_speech_token=llm_prompt_speech_token.to(self.device),
-                                            prompt_speech_token_len=torch.tensor([llm_prompt_speech_token.shape[1]], dtype=torch.int32).to(self.device),
-                                            embedding=llm_embedding.to(self.device),
-                                            uuid=uuid,
-                                            progress_callback=progress_callback):
-                    self.tts_speech_token_dict[uuid].append(i)
-        self.llm_end_dict[uuid] = True
+        # TTS Audio Suite patch: capture worker failures for _raise_if_llm_failed().
+        try:
+            with self.llm_context, self._llm_autocast_context():
+                if isinstance(text, Generator):
+                    assert (self.__class__.__name__ != 'CosyVoiceModel') and not hasattr(self.llm, 'vllm'), 'streaming input text is only implemented for CosyVoice2/3 and do not support vllm!'
+                    for i in self.llm.inference_bistream(text=text,
+                                                         prompt_text=prompt_text.to(self.device),
+                                                         prompt_text_len=torch.tensor([prompt_text.shape[1]], dtype=torch.int32).to(self.device),
+                                                         prompt_speech_token=llm_prompt_speech_token.to(self.device),
+                                                         prompt_speech_token_len=torch.tensor([llm_prompt_speech_token.shape[1]], dtype=torch.int32).to(self.device),
+                                                         embedding=llm_embedding.to(self.device)):
+                        self.tts_speech_token_dict[uuid].append(i)
+                else:
+                    for i in self.llm.inference(text=text.to(self.device),
+                                                text_len=torch.tensor([text.shape[1]], dtype=torch.int32).to(self.device),
+                                                prompt_text=prompt_text.to(self.device),
+                                                prompt_text_len=torch.tensor([prompt_text.shape[1]], dtype=torch.int32).to(self.device),
+                                                prompt_speech_token=llm_prompt_speech_token.to(self.device),
+                                                prompt_speech_token_len=torch.tensor([llm_prompt_speech_token.shape[1]], dtype=torch.int32).to(self.device),
+                                                embedding=llm_embedding.to(self.device),
+                                                uuid=uuid,
+                                                progress_callback=progress_callback):
+                        self.tts_speech_token_dict[uuid].append(i)
+        except Exception as error:
+            self.llm_error_dict[uuid] = error
+        finally:
+            self.llm_end_dict[uuid] = True
 
     def vc_job(self, source_speech_token, uuid):
         self.tts_speech_token_dict[uuid] = source_speech_token.flatten().tolist()
@@ -188,6 +222,7 @@ class CosyVoiceModel:
         this_uuid = str(uuid.uuid1())
         with self.lock:
             self.tts_speech_token_dict[this_uuid], self.llm_end_dict[this_uuid] = [], False
+            self.llm_error_dict[this_uuid] = None
             self.hift_cache_dict[this_uuid] = None
             self.mel_overlap_dict[this_uuid] = torch.zeros(1, 80, 0)
             self.flow_cache_dict[this_uuid] = torch.zeros(1, 80, 0, 2)
@@ -217,6 +252,7 @@ class CosyVoiceModel:
                 if self.llm_end_dict[this_uuid] is True and len(self.tts_speech_token_dict[this_uuid]) < token_hop_len + self.token_overlap_len:
                     break
             p.join()
+            self._raise_if_llm_failed(this_uuid)
             # deal with remain tokens, make sure inference remain token len equals token_hop_len when cache_speech is not None
             this_tts_speech_token = torch.tensor(self.tts_speech_token_dict[this_uuid]).unsqueeze(dim=0)
             this_tts_speech = self.token2wav(token=this_tts_speech_token,
@@ -229,6 +265,7 @@ class CosyVoiceModel:
         else:
             # deal with all tokens
             p.join()
+            self._raise_if_llm_failed(this_uuid)
             this_tts_speech_token = torch.tensor(self.tts_speech_token_dict[this_uuid]).unsqueeze(dim=0)
             this_tts_speech = self.token2wav(token=this_tts_speech_token,
                                              prompt_token=flow_prompt_speech_token,
@@ -241,6 +278,7 @@ class CosyVoiceModel:
         with self.lock:
             self.tts_speech_token_dict.pop(this_uuid)
             self.llm_end_dict.pop(this_uuid)
+            self.llm_error_dict.pop(this_uuid, None)
             self.mel_overlap_dict.pop(this_uuid)
             self.hift_cache_dict.pop(this_uuid)
             self.flow_cache_dict.pop(this_uuid)
@@ -274,6 +312,7 @@ class CosyVoice2Model(CosyVoiceModel):
         # dict used to store session related variable
         self.tts_speech_token_dict = {}
         self.llm_end_dict = {}
+        self.llm_error_dict = {}
         self.hift_cache_dict = {}
 
     def load_jit(self, flow_encoder_model):
@@ -336,6 +375,7 @@ class CosyVoice2Model(CosyVoiceModel):
         this_uuid = str(uuid.uuid1())
         with self.lock:
             self.tts_speech_token_dict[this_uuid], self.llm_end_dict[this_uuid] = [], False
+            self.llm_error_dict[this_uuid] = None
             self.hift_cache_dict[this_uuid] = None
         if source_speech_token.shape[1] == 0:
             p = threading.Thread(target=self.llm_job, args=(text, prompt_text, llm_prompt_speech_token, llm_embedding, this_uuid, progress_callback))
@@ -363,6 +403,7 @@ class CosyVoice2Model(CosyVoiceModel):
                 if self.llm_end_dict[this_uuid] is True and len(self.tts_speech_token_dict[this_uuid]) - token_offset < this_token_hop_len + self.flow.pre_lookahead_len:
                     break
             p.join()
+            self._raise_if_llm_failed(this_uuid)
             # deal with remain tokens, make sure inference remain token len equals token_hop_len when cache_speech is not None
             this_tts_speech_token = torch.tensor(self.tts_speech_token_dict[this_uuid]).unsqueeze(dim=0)
             this_tts_speech = self.token2wav(token=this_tts_speech_token,
@@ -376,6 +417,7 @@ class CosyVoice2Model(CosyVoiceModel):
         else:
             # deal with all tokens
             p.join()
+            self._raise_if_llm_failed(this_uuid)
             this_tts_speech_token = torch.tensor(self.tts_speech_token_dict[this_uuid]).unsqueeze(dim=0)
             this_tts_speech = self.token2wav(token=this_tts_speech_token,
                                              prompt_token=flow_prompt_speech_token,
@@ -389,6 +431,7 @@ class CosyVoice2Model(CosyVoiceModel):
         with self.lock:
             self.tts_speech_token_dict.pop(this_uuid)
             self.llm_end_dict.pop(this_uuid)
+            self.llm_error_dict.pop(this_uuid, None)
             self.hift_cache_dict.pop(this_uuid)
         if torch.cuda.is_available():
             torch.cuda.empty_cache()
@@ -406,7 +449,9 @@ class CosyVoice3Model(CosyVoice2Model):
         self.llm = llm
         self.flow = flow
         self.hift = hift
-        self.fp16 = fp16
+        # TTS Audio Suite patch: enable BF16 LLM autocast on ROCm only.
+        self._rocm_bf16_llm_autocast = torch.cuda.is_available() and bool(torch.version.hip)
+        self.fp16 = fp16 and not self._rocm_bf16_llm_autocast
         # NOTE must matching training static_chunk_size
         self.token_hop_len = 25
         # rtf and decoding related
@@ -415,6 +460,7 @@ class CosyVoice3Model(CosyVoice2Model):
         # dict used to store session related variable
         self.tts_speech_token_dict = {}
         self.llm_end_dict = {}
+        self.llm_error_dict = {}
         self.hift_cache_dict = {}
 
     def token2wav(self, token, prompt_token, prompt_feat, embedding, token_offset, uuid, stream=False, finalize=False, speed=1.0):

engines/dots_tts/__init__.py

@@ -0,0 +1,6 @@
+"""Dots TTS engine integration."""
+
+from .dots_tts_engine import DotsTTSEngine
+from .dots_tts_downloader import DotsTTSDownloader
+
+__all__ = ["DotsTTSEngine", "DotsTTSDownloader"]

engines/dots_tts/dots_tts_downloader.py

@@ -0,0 +1,143 @@
+"""
+Dots TTS model downloader.
+
+Downloads official Rednote dots.tts checkpoints into
+ComfyUI/models/TTS/dots_tts/ instead of hidden Hugging Face cache folders.
+"""
+
+import os
+from typing import Dict, Optional
+
+import folder_paths
+from huggingface_hub import snapshot_download
+
+from utils.hf_download_logging import quiet_hf_download_logs
+from utils.models.extra_paths import get_all_tts_model_paths, get_preferred_download_path
+
+
+class DotsTTSDownloader:
+    """Resolve and download official dots.tts model folders."""
+
+    MODELS: Dict[str, Dict[str, str]] = {
+        "dots.tts-base": {
+            "repo_id": "rednote-hilab/dots.tts-base",
+            "description": "Official Dots TTS base checkpoint",
+        },
+        "dots.tts-soar": {
+            "repo_id": "rednote-hilab/dots.tts-soar",
+            "description": "Official Dots TTS SOAR checkpoint",
+        },
+        "dots.tts-mf": {
+            "repo_id": "rednote-hilab/dots.tts-mf",
+            "description": "Official Dots TTS MeanFlow checkpoint",
+        },
+    }
+
+    REQUIRED_FILES = [
+        "added_tokens.json",
+        "chat_template.jinja",
+        "config.json",
+        "latent_stats.pt",
+        "llm_config.json",
+        "merges.txt",
+        "model.safetensors",
+        "speaker_encoder.safetensors",
+        "special_tokens_map.json",
+        "tokenizer.json",
+        "tokenizer_config.json",
+        "vocab.json",
+        "vocoder.safetensors",
+    ]
+
+    def __init__(self, base_path: Optional[str] = None):
+        if base_path is None:
+            try:
+                self.base_path = get_preferred_download_path(model_type="TTS", engine_name="dots_tts")
+            except Exception:
+                self.base_path = os.path.join(folder_paths.models_dir, "TTS", "dots_tts")
+        else:
+            self.base_path = base_path
+        os.makedirs(self.base_path, exist_ok=True)
+
+    def resolve_model_path(self, model_identifier: str) -> str:
+        """Resolve local: names, absolute paths, or known official model identifiers."""
+        if not model_identifier:
+            model_identifier = "dots.tts-soar"
+
+        if os.path.isabs(model_identifier) and os.path.exists(model_identifier):
+            return model_identifier
+
+        if os.path.isdir(model_identifier):
+            return model_identifier
+
+        if model_identifier.startswith("local:"):
+            local_name = model_identifier[6:]
+            for base_path in get_all_tts_model_paths("TTS"):
+                for folder_name in ("dots_tts", "DOTS", ""):
+                    candidate = os.path.join(base_path, folder_name, local_name) if folder_name else os.path.join(base_path, local_name)
+                    if os.path.exists(candidate):
+                        print(f"📁 Using local Dots TTS model: {candidate}")
+                        return candidate
+            raise FileNotFoundError(f"Local Dots TTS model not found: {local_name}")
+
+        if model_identifier in self.MODELS:
+            return self.get_model_path(model_identifier)
+
+        raise FileNotFoundError(f"Unknown Dots TTS model identifier: {model_identifier}")
+
+    def get_model_path(self, model_name: str) -> str:
+        model_dir = os.path.join(self.base_path, model_name)
+        if not os.path.exists(model_dir) or not self._is_model_complete(model_dir):
+            return self.download_model(model_name, force=os.path.exists(model_dir))
+        return model_dir
+
+    def download_model(self, model_name: str, force: bool = False) -> str:
+        if model_name not in self.MODELS:
+            raise ValueError(f"Unknown Dots TTS model: {model_name}")
+
+        model_info = self.MODELS[model_name]
+        repo_id = model_info["repo_id"]
+        model_dir = os.path.join(self.base_path, model_name)
+
+        print(f"\n{'=' * 60}")
+        print("📦 Dots TTS Model Download")
+        print(f"{'=' * 60}")
+        print(f"Model: {model_name}")
+        print(f"Description: {model_info['description']}")
+        print(f"Repository: {repo_id}")
+        print(f"Target: {model_dir}")
+        print(f"{'=' * 60}\n")
+
+        try:
+            with quiet_hf_download_logs():
+                snapshot_download(
+                    repo_id=repo_id,
+                    local_dir=model_dir,
+                    local_dir_use_symlinks=False,
+                    resume_download=True,
+                    force_download=force,
+                )
+        except Exception as e:
+            raise RuntimeError(f"Failed to download {model_name} from {repo_id}: {e}") from e
+
+        if not self._is_model_complete(model_dir):
+            raise RuntimeError(f"Downloaded Dots TTS model is incomplete: {model_dir}")
+
+        print(f"✅ Dots TTS model ready: {model_dir}")
+        return model_dir
+
+    def _is_model_complete(self, model_dir: str) -> bool:
+        if not os.path.isdir(model_dir):
+            return False
+
+        missing = [
+            rel_path
+            for rel_path in self.REQUIRED_FILES
+            if not os.path.exists(os.path.join(model_dir, rel_path))
+        ]
+        if missing:
+            print(f"❌ Dots TTS model incomplete. Missing {len(missing)} file(s):")
+            for rel_path in missing:
+                print(f"   - {rel_path}")
+            return False
+        return True

engines/dots_tts/dots_tts_engine.py

@@ -0,0 +1,475 @@
+"""
+Dots TTS engine wrapper.
+
+Wraps the official dots.tts runtime with ComfyUI-friendly model lifecycle hooks.
+"""
+
+import os
+import sys
+import types
+import importlib
+import logging
+import time
+import warnings
+import gc
+from contextlib import contextmanager
+from typing import Any, Dict, Optional
+
+import torch
+
+current_dir = os.path.dirname(__file__)
+engines_dir = os.path.dirname(current_dir)
+project_root = os.path.dirname(engines_dir)
+if project_root not in sys.path:
+    sys.path.insert(0, project_root)
+
+from engines.dots_tts.dots_tts_downloader import DotsTTSDownloader
+from engines.dots_tts.progress_callback import DotsTTSProgressTracker
+from utils.device import resolve_torch_device
+
+
+class _SuppressDotsModelTypeWarning(logging.Filter):
+    """Drop the upstream bogus model-type warning while keeping real ones."""
+
+    def filter(self, record: logging.LogRecord) -> bool:
+        try:
+            message = record.getMessage()
+        except Exception:
+            return True
+        return "model of type `dots_tts` to instantiate a model of type" not in message
+
+
+class DotsTTSEngine:
+    """Thin wrapper around the official DotsTtsRuntime."""
+
+    SAMPLE_RATE = 48000
+    _normalizer_warning_shown = False
+
+    def __init__(
+        self,
+        model_name: str = "dots.tts-soar",
+        device: str = "auto",
+        precision: str = "auto",
+        optimize: bool = False,
+        max_generate_length: int = 500,
+        model_dir: Optional[str] = None,
+    ):
+        self.model_name = model_name
+        self.device = resolve_torch_device(device)
+        self.precision = self._resolve_precision(precision, self.device)
+        self.optimize = bool(optimize)
+        self.max_generate_length = int(max_generate_length)
+        self.model_dir = model_dir or DotsTTSDownloader().resolve_model_path(model_name)
+        self._runtime = None
+
+    @staticmethod
+    def _resolve_precision(precision: str, device: str) -> str:
+        normalized = str(precision or "auto").lower()
+        if normalized in {"bfloat16", "float16", "float32"}:
+            if device == "cpu" and normalized != "float32":
+                return "float32"
+            return normalized
+
+        if device == "cpu":
+            return "float32"
+
+        if torch.cuda.is_available():
+            major, _minor = torch.cuda.get_device_capability()
+            if major >= 8:
+                return "bfloat16"
+        return "float16"
+
+    def _import_runtime(self):
+        importlib.invalidate_caches()
+
+        nodes_dir = os.path.join(project_root, "nodes")
+        blocked_paths = {
+            os.path.abspath(engines_dir),
+            os.path.abspath(nodes_dir),
+        }
+        original_sys_path = list(sys.path)
+
+        # ComfyUI can place the suite's `engines/` directory directly on sys.path.
+        # That makes our local `engines/dots_tts` package shadow the pip-installed
+        # official `dots_tts` package we actually need here.
+        sys.path[:] = [
+            path for path in sys.path
+            if os.path.abspath(path or os.getcwd()) not in blocked_paths
+        ]
+
+        stale_modules = []
+        for module_name, module in list(sys.modules.items()):
+            if not (module_name == "dots_tts" or module_name.startswith("dots_tts.")):
+                continue
+            module_file = getattr(module, "__file__", "") or ""
+            if module_file and os.path.abspath(module_file).startswith(os.path.abspath(project_root)):
+                stale_modules.append((module_name, module))
+                del sys.modules[module_name]
+        try:
+            with self._text_normalizer_compat():
+                from dots_tts.runtime import DotsTtsRuntime
+        except Exception as e:
+            for module_name, module in stale_modules:
+                sys.modules.setdefault(module_name, module)
+            raise ImportError(
+                "Failed to import the official `dots_tts` package in the active ComfyUI Python "
+                "environment. This is usually either a missing install or an import-path "
+                f"collision. Active Python: {sys.executable}"
+            ) from e
+        finally:
+            sys.path[:] = original_sys_path
+        return DotsTtsRuntime
+
+    @staticmethod
+    @contextmanager
+    def _suppress_noisy_runtime_logs():
+        """Mute verbose Dots chatter while keeping actionable Transformers warnings."""
+        dots_logger = None
+        dots_logger_disabled = False
+        transformers_filter = _SuppressDotsModelTypeWarning()
+        transformer_loggers = []
+
+        try:
+            from loguru import logger as dots_logger
+
+            dots_logger.disable("dots_tts")
+            dots_logger_disabled = True
+
+            transformer_loggers = [
+                logging.getLogger("transformers"),
+                logging.getLogger("transformers.configuration_utils"),
+                logging.getLogger("transformers.modeling_utils"),
+                logging.getLogger("transformers.tokenization_utils_base"),
+            ]
+            for logger in transformer_loggers:
+                logger.addFilter(transformers_filter)
+        except Exception:
+            pass
+
+        try:
+            with warnings.catch_warnings():
+                warnings.filterwarnings(
+                    "ignore",
+                    message=r".*model of type `dots_tts` to instantiate a model of type.*",
+                )
+                yield
+        finally:
+            if dots_logger is not None and dots_logger_disabled:
+                try:
+                    dots_logger.enable("dots_tts")
+                except Exception:
+                    pass
+
+            for logger in transformer_loggers:
+                try:
+                    logger.removeFilter(transformers_filter)
+                except Exception:
+                    pass
+
+    @staticmethod
+    @contextmanager
+    def _patch_tokenizer_regex_fix():
+        """Force the upstream Dots tokenizer load to opt into the Mistral regex fix."""
+        try:
+            from transformers import AutoTokenizer
+        except Exception:
+            yield
+            return
+
+        original_from_pretrained = AutoTokenizer.from_pretrained
+
+        def patched_from_pretrained(*args, **kwargs):
+            kwargs.setdefault("fix_mistral_regex", True)
+            return original_from_pretrained(*args, **kwargs)
+
+        AutoTokenizer.from_pretrained = patched_from_pretrained
+        try:
+            yield
+        finally:
+            AutoTokenizer.from_pretrained = original_from_pretrained
+
+    @staticmethod
+    @contextmanager
+    def _text_normalizer_compat():
+        """Temporarily provide Dots' expected tn package when unavailable."""
+        normalizers_available = False
+        try:
+            from tn.chinese.normalizer import Normalizer as _ZhNormalizer  # noqa: F401
+            from tn.english.normalizer import Normalizer as _EnNormalizer  # noqa: F401
+            normalizers_available = True
+        except Exception:
+            pass
+
+        if normalizers_available:
+            yield
+            return
+
+        class _NoOpNormalizer:
+            def normalize(self, text: str) -> str:
+                return text
+
+        tn_module = types.ModuleType("tn")
+        chinese_module = types.ModuleType("tn.chinese")
+        chinese_normalizer_module = types.ModuleType("tn.chinese.normalizer")
+        english_module = types.ModuleType("tn.english")
+        english_normalizer_module = types.ModuleType("tn.english.normalizer")
+
+        # Mark package modules as packages so nested imports work even when an
+        # unrelated top-level module named `tn` is already installed.
+        tn_module.__path__ = []
+        chinese_module.__path__ = []
+        english_module.__path__ = []
+
+        chinese_normalizer_module.Normalizer = _NoOpNormalizer
+        english_normalizer_module.Normalizer = _NoOpNormalizer
+
+        chinese_module.normalizer = chinese_normalizer_module
+        english_module.normalizer = english_normalizer_module
+        tn_module.chinese = chinese_module
+        tn_module.english = english_module
+
+        fallback_modules = {
+            "tn": tn_module,
+            "tn.chinese": chinese_module,
+            "tn.chinese.normalizer": chinese_normalizer_module,
+            "tn.english": english_module,
+            "tn.english.normalizer": english_normalizer_module,
+        }
+        missing = object()
+        previous_modules = {
+            name: sys.modules.get(name, missing)
+            for name in fallback_modules
+        }
+        sys.modules.update(fallback_modules)
+        try:
+            if not DotsTTSEngine._normalizer_warning_shown:
+                print("[Dots TTS] WeTextProcessing not available; normalize_text will use a no-op fallback")
+                DotsTTSEngine._normalizer_warning_shown = True
+            yield
+        finally:
+            for name, previous in previous_modules.items():
+                if previous is missing:
+                    sys.modules.pop(name, None)
+                else:
+                    sys.modules[name] = previous
+
+    def _ensure_runtime_loaded(self):
+        if self._runtime is not None:
+            return
+
+        DotsTtsRuntime = self._import_runtime()
+        with self._patch_tokenizer_regex_fix(), self._suppress_noisy_runtime_logs():
+            self._runtime = DotsTtsRuntime.from_pretrained(
+                self.model_dir,
+                precision=self.precision,
+                optimize=self.optimize,
+                max_generate_length=self.max_generate_length,
+            )
+        self._ensure_runtime_device()
+
+    def _current_runtime_device(self) -> Optional[str]:
+        if self._runtime is None:
+            return None
+        try:
+            first_param = next(self._runtime.model.parameters())
+            return str(first_param.device)
+        except Exception:
+            return str(getattr(self._runtime, "device", self.device))
+
+    def _move_runtime_to(self, target_device: str):
+        if self._runtime is None:
+            return
+
+        device_obj = torch.device(target_device)
+        runtime = self._runtime
+        runtime.model = runtime.model.to(device_obj).eval()
+        runtime.device = device_obj
+
+    def _ensure_runtime_device(self):
+        if self._runtime is None:
+            return
+        target_device = resolve_torch_device(self.device)
+        current_device = self._current_runtime_device()
+        if self._devices_equivalent(current_device, target_device):
+            return
+        if current_device != target_device:
+            print(f"🔄 Dots TTS: moving runtime from {current_device} to {target_device}")
+            self._move_runtime_to(target_device)
+
+    @staticmethod
+    def _devices_equivalent(current_device: Optional[str], target_device: Optional[str]) -> bool:
+        if current_device == target_device:
+            return True
+        if not current_device or not target_device:
+            return False
+        current = str(current_device)
+        target = str(target_device)
+        if current.startswith("cuda") and target.startswith("cuda"):
+            return current in {"cuda", "cuda:0"} and target in {"cuda", "cuda:0"}
+        return False
+
+    @staticmethod
+    def _clear_runtime_caches(runtime: Any) -> None:
+        if runtime is None:
+            return
+
+        model = getattr(runtime, "model", None)
+        if model is not None:
+            try:
+                if hasattr(model, "set_optimize"):
+                    model.set_optimize(False)
+            except Exception:
+                pass
+
+            for attr_name in (
+                "_compiled_models",
+                "_prompt_feature_cache",
+                "_static_generate_workspaces",
+                "_fm_decode_workspaces",
+            ):
+                try:
+                    cache = getattr(model, attr_name, None)
+                    if hasattr(cache, "clear"):
+                        cache.clear()
+                except Exception:
+                    pass
+
+        for attr_name in (
+            "_double_streaming_prompt_g_cond_cache",
+            "_double_streaming_silence_audio_patch_cache",
+        ):
+            try:
+                cache = getattr(runtime, attr_name, None)
+                if hasattr(cache, "clear"):
+                    cache.clear()
+            except Exception:
+                pass
+
+    def unload_runtime(self):
+        """Fully tear down the Dots runtime so Clear VRAM actually releases it."""
+        runtime = self._runtime
+        if runtime is None:
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+                if hasattr(torch.cuda, "ipc_collect"):
+                    try:
+                        torch.cuda.ipc_collect()
+                    except Exception:
+                        pass
+            gc.collect()
+            return
+
+        self._clear_runtime_caches(runtime)
+
+        model = getattr(runtime, "model", None)
+        try:
+            if model is not None and hasattr(model, "to"):
+                model.to("cpu")
+        except Exception:
+            pass
+
+        try:
+            runtime.model = None
+        except Exception:
+            pass
+
+        self._runtime = None
+
+        try:
+            import torch._dynamo as torch_dynamo
+
+            torch_dynamo.reset()
+        except Exception:
+            pass
+
+        del model
+        del runtime
+        gc.collect()
+
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            if hasattr(torch.cuda, "ipc_collect"):
+                try:
+                    torch.cuda.ipc_collect()
+                except Exception:
+                    pass
+
+    def to(self, device):
+        """ComfyUI unload/reload hook."""
+        self.device = str(device) if isinstance(device, str) else str(torch.device(device))
+        self._ensure_runtime_device()
+        return self
+
+    def unload(self):
+        """Explicit unload hook used by engine-specific ComfyUI handlers."""
+        self.unload_runtime()
+
+    def generate(
+        self,
+        text: str,
+        prompt_audio_path: Optional[str] = None,
+        prompt_text: Optional[str] = None,
+        language: Optional[str] = None,
+        num_steps: int = 10,
+        guidance_scale: float = 1.2,
+        speaker_scale: float = 1.5,
+        normalize_text: bool = False,
+        template_name: Optional[str] = "tts",
+        seed: int = 42,
+    ) -> Dict[str, Any]:
+        self._ensure_runtime_loaded()
+        self._ensure_runtime_device()
+
+        seed_value = int(seed or 0)
+        if seed_value > 0:
+            torch.manual_seed(seed_value)
+            if torch.cuda.is_available():
+                torch.cuda.manual_seed_all(seed_value)
+
+        chunks = []
+        tracker = DotsTTSProgressTracker(
+            max_generate_length=self.max_generate_length,
+            sample_rate=self.SAMPLE_RATE,
+        )
+        start_time = time.time()
+
+        with self._patch_tokenizer_regex_fix(), self._suppress_noisy_runtime_logs():
+            stream = self._runtime.generate_stream(
+                text=text,
+                prompt_audio_path=prompt_audio_path,
+                prompt_text=prompt_text,
+                language=language,
+                template_name=template_name,
+                num_steps=int(num_steps),
+                guidance_scale=float(guidance_scale),
+                speaker_scale=float(speaker_scale),
+                normalize_text=bool(normalize_text),
+            )
+            for chunk in stream:
+                if not isinstance(chunk, torch.Tensor):
+                    chunk = torch.tensor(chunk, dtype=torch.float32)
+                chunk = chunk.detach().float().cpu()
+                if chunk.dim() > 1:
+                    chunk = chunk.squeeze()
+                chunks.append(chunk)
+                tracker.update(chunk.shape[-1] if chunk.numel() > 0 else 0)
+
+        tracker.end()
+
+        if chunks:
+            audio = torch.cat(chunks, dim=-1)
+        else:
+            audio = torch.zeros(0, dtype=torch.float32)
+
+        time_used = time.time() - start_time
+        duration_seconds = audio.shape[-1] / self.SAMPLE_RATE if audio.numel() > 0 else 0.0
+        rtf = time_used / duration_seconds if duration_seconds > 0 else float("inf")
+
+        return {
+            "fid": None,
+            "audio": audio,
+            "sample_rate": self.SAMPLE_RATE,
+            "time_used": time_used,
+            "rtf": rtf,
+            "profiling": None,
+        }

engines/dots_tts/languages.py

@@ -0,0 +1,137 @@
+"""
+Language helpers for Dots TTS.
+"""
+
+from typing import Optional
+
+
+DOTS_LANGUAGE_CODE_BY_NAME = {
+    "arabic": "AR",
+    "czech": "CS",
+    "german": "DE",
+    "greek": "EL",
+    "english": "EN",
+    "spanish": "ES",
+    "finnish": "FI",
+    "french": "FR",
+    "hindi": "HI",
+    "indonesian": "ID",
+    "italian": "IT",
+    "japanese": "JA",
+    "korean": "KO",
+    "dutch": "NL",
+    "polish": "PL",
+    "portuguese": "PT",
+    "pt": "PT",
+    "pt-br": "PT",
+    "pt-pt": "PT",
+    "po": "PT",
+    "romanian": "RO",
+    "russian": "RU",
+    "thai": "TH",
+    "turkish": "TR",
+    "ukrainian": "UK",
+    "uk": "UK",
+    "vietnamese": "VI",
+    "cantonese": "YUE",
+    "yue": "YUE",
+    "chinese": "ZH",
+    "mandarin": "ZH",
+    "zh": "ZH",
+    "zh-cn": "ZH",
+    "zh-tw": "ZH",
+}
+
+DOTS_LANGUAGE_DISPLAY_BY_CODE = {
+    "AR": "Arabic",
+    "CS": "Czech",
+    "DE": "German",
+    "EL": "Greek",
+    "EN": "English",
+    "ES": "Spanish",
+    "FI": "Finnish",
+    "FR": "French",
+    "HI": "Hindi",
+    "ID": "Indonesian",
+    "IT": "Italian",
+    "JA": "Japanese",
+    "KO": "Korean",
+    "NL": "Dutch",
+    "PL": "Polish",
+    "PT": "Portuguese",
+    "RO": "Romanian",
+    "RU": "Russian",
+    "TH": "Thai",
+    "TR": "Turkish",
+    "UK": "Ukrainian",
+    "VI": "Vietnamese",
+    "YUE": "Cantonese",
+    "ZH": "Chinese",
+}
+
+DOTS_LANGUAGE_OPTIONS = [
+    "Auto",
+    "None",
+    "Arabic",
+    "Czech",
+    "German",
+    "Greek",
+    "English",
+    "Spanish",
+    "Finnish",
+    "French",
+    "Hindi",
+    "Indonesian",
+    "Italian",
+    "Japanese",
+    "Korean",
+    "Dutch",
+    "Polish",
+    "Portuguese",
+    "Romanian",
+    "Russian",
+    "Thai",
+    "Turkish",
+    "Ukrainian",
+    "Vietnamese",
+    "Cantonese",
+    "Chinese",
+]
+
+
+def normalize_dots_language(language: Optional[str]) -> Optional[str]:
+    if language is None:
+        return None
+
+    normalized = str(language).strip()
+    if not normalized:
+        return None
+
+    lowered = normalized.lower()
+    if lowered in {"auto", "auto detect", "auto_detect"}:
+        return "auto_detect"
+    if lowered in {"none", "off", "no tag", "no language tag"}:
+        return None
+
+    mapped = DOTS_LANGUAGE_CODE_BY_NAME.get(lowered)
+    if mapped:
+        return mapped
+
+    upper_value = normalized.upper()
+    if upper_value in DOTS_LANGUAGE_DISPLAY_BY_CODE:
+        return upper_value
+
+    if lowered == "yue":
+        return "YUE"
+
+    return upper_value
+
+
+def format_dots_language_display(language: Optional[str]) -> str:
+    normalized = normalize_dots_language(language)
+    if normalized == "auto_detect":
+        return "Auto"
+    if normalized is None:
+        return "None"
+
+    return DOTS_LANGUAGE_DISPLAY_BY_CODE.get(normalized, str(language).strip() or "Auto")

engines/dots_tts/progress_callback.py

@@ -0,0 +1,66 @@
+"""
+Dots TTS progress callback.
+
+Provides suite-owned console progress output during streaming generation.
+"""
+
+import time
+
+
+class DotsTTSProgressTracker:
+    """Track Dots streaming progress with live it/s and RTF reporting."""
+
+    def __init__(self, max_generate_length: int, sample_rate: int):
+        self.max_generate_length = max(1, int(max_generate_length or 1))
+        self.sample_rate = max(1, int(sample_rate or 1))
+        self.generated_chunks = 0
+        self.emitted_samples = 0
+        self.start_time = time.time()
+        self.last_print_time = self.start_time
+        self.last_print_chunks = 0
+
+    def update(self, chunk_samples: int) -> None:
+        self.generated_chunks += 1
+        self.emitted_samples += max(0, int(chunk_samples or 0))
+
+        current_time = time.time()
+        should_print = (
+            self.generated_chunks == 1 or
+            current_time - self.last_print_time >= 0.5
+        )
+        if not should_print:
+            return
+
+        chunk_delta = self.generated_chunks - self.last_print_chunks
+        time_delta = current_time - self.last_print_time
+        its = chunk_delta / time_delta if time_delta > 0 else 0.0
+        elapsed = current_time - self.start_time
+        audio_seconds = self.emitted_samples / self.sample_rate
+        rtf = elapsed / audio_seconds if audio_seconds > 0 else 0.0
+
+        bar_width = 12
+        display_chunks = min(self.generated_chunks, self.max_generate_length)
+        filled = int(bar_width * display_chunks / self.max_generate_length)
+        filled = min(filled, bar_width)
+        progress_bar = f"[{'█' * filled}{'░' * (bar_width - filled)}] {display_chunks}/{self.max_generate_length}"
+
+        print(
+            f"\r   Progress: {progress_bar} | {its:.1f} it/s | {elapsed:.0f}s | "
+            f"{audio_seconds:.1f}s audio | RTF {rtf:.2f}      ",
+            end="",
+            flush=True,
+        )
+
+        self.last_print_time = current_time
+        self.last_print_chunks = self.generated_chunks
+
+    def end(self) -> None:
+        total_time = time.time() - self.start_time
+        avg_its = self.generated_chunks / total_time if total_time > 0 else 0.0
+        audio_seconds = self.emitted_samples / self.sample_rate
+        rtf = total_time / audio_seconds if audio_seconds > 0 else 0.0
+        print(
+            f"\r   Complete: {self.generated_chunks} chunks in {total_time:.1f}s "
+            f"(avg {avg_its:.1f} it/s, {audio_seconds:.1f}s audio, RTF {rtf:.2f})"
+            + " " * 20
+        )

engines/dramabox/__init__.py

@@ -0,0 +1,6 @@
+"""DramaBox engine integration."""
+
+from .dramabox_downloader import DramaBoxDownloader
+from .dramabox_engine import DramaBoxEngine
+
+__all__ = ["DramaBoxDownloader", "DramaBoxEngine"]

engines/dramabox/dramabox_downloader.py

@@ -0,0 +1,150 @@
+"""Organized model download and discovery for official DramaBox."""
+
+import os
+from typing import Dict, List, Optional
+
+import folder_paths
+
+from utils.downloads.unified_downloader import unified_downloader
+from utils.models.extra_paths import get_all_tts_model_paths, get_preferred_download_path
+
+
+class DramaBoxDownloader:
+    """Resolve DramaBox checkpoints without using Hugging Face cache storage."""
+
+    MODEL_NAME = "DramaBox"
+    DRAMABOX_REPO = "ResembleAI/Dramabox"
+    GEMMA_REPO = "unsloth/gemma-3-12b-it-bnb-4bit"
+
+    DRAMABOX_FILES = [
+        "dramabox-dit-v1.safetensors",
+        "dramabox-audio-components.safetensors",
+        "assets/silence_latent_frame.pt",
+    ]
+    GEMMA_FILES = [
+        "config.json",
+        "generation_config.json",
+        "model-00001-of-00002.safetensors",
+        "model-00002-of-00002.safetensors",
+        "model.safetensors.index.json",
+        "tokenizer.json",
+        "tokenizer.model",
+        "tokenizer_config.json",
+        "special_tokens_map.json",
+        "added_tokens.json",
+        "preprocessor_config.json",
+        "processor_config.json",
+        "chat_template.jinja",
+        "chat_template.json",
+    ]
+
+    def __init__(self, base_path: Optional[str] = None):
+        if base_path is None:
+            try:
+                self.base_path = get_preferred_download_path(
+                    model_type="TTS", engine_name="dramabox"
+                )
+            except Exception:
+                self.base_path = os.path.join(folder_paths.models_dir, "TTS", "dramabox")
+        else:
+            self.base_path = base_path
+        os.makedirs(self.base_path, exist_ok=True)
+
+    def get_available_models(self) -> List[str]:
+        models = [self.MODEL_NAME]
+        for base_path in get_all_tts_model_paths("TTS"):
+            for folder_name in ("dramabox", "DramaBox"):
+                root = os.path.join(base_path, folder_name)
+                if not os.path.isdir(root):
+                    continue
+                for item in sorted(os.listdir(root)):
+                    candidate = os.path.join(root, item)
+                    if os.path.isdir(candidate) and self._is_model_complete(candidate):
+                        local_name = f"local:{item}"
+                        if local_name not in models:
+                            models.insert(0, local_name)
+        return models
+
+    def resolve_model_path(self, model_identifier: str = MODEL_NAME) -> Dict[str, str]:
+        model_identifier = model_identifier or self.MODEL_NAME
+        if os.path.isabs(model_identifier) and os.path.isdir(model_identifier):
+            return self._paths_for(model_identifier)
+
+        if model_identifier.startswith("local:"):
+            local_name = model_identifier[6:]
+            for base_path in get_all_tts_model_paths("TTS"):
+                for folder_name in ("dramabox", "DramaBox"):
+                    candidate = os.path.join(base_path, folder_name, local_name)
+                    if self._is_model_complete(candidate):
+                        print(f"📁 Using local DramaBox model: {candidate}")
+                        return self._paths_for(candidate)
+            raise FileNotFoundError(f"Local DramaBox model not found or incomplete: {local_name}")
+
+        if model_identifier != self.MODEL_NAME:
+            raise ValueError(f"Unknown DramaBox model: {model_identifier}")
+
+        model_dir = os.path.join(self.base_path, self.MODEL_NAME)
+        if not self._is_model_complete(model_dir):
+            self.download_model(model_dir)
+        return self._paths_for(model_dir)
+
+    def download_model(self, model_dir: Optional[str] = None) -> str:
+        model_dir = model_dir or os.path.join(self.base_path, self.MODEL_NAME)
+        gemma_dir = os.path.join(model_dir, "gemma-3-12b-it-bnb-4bit")
+
+        print("\n" + "=" * 60)
+        print("📦 DramaBox Model Download")
+        print("=" * 60)
+        print(f"DramaBox: {self.DRAMABOX_REPO}")
+        print(f"Gemma encoder: {self.GEMMA_REPO}")
+        print(f"Target: {model_dir}")
+        print("License: LTX-2 Community License (commercial threshold applies)")
+        print("=" * 60 + "\n")
+
+        base_files = [
+            {"remote": rel_path, "local": rel_path}
+            for rel_path in self.DRAMABOX_FILES
+        ]
+        result = unified_downloader.download_huggingface_model(
+            repo_id=self.DRAMABOX_REPO,
+            model_name=self.MODEL_NAME,
+            files=base_files,
+            engine_type="dramabox",
+            target_dir=model_dir,
+        )
+        if not result:
+            raise RuntimeError("Failed to download official DramaBox weights")
+
+        unified_downloader.download_huggingface_snapshot(
+            repo_id=self.GEMMA_REPO,
+            target_dir=gemma_dir,
+            allow_patterns=self.GEMMA_FILES,
+            required_files=self.GEMMA_FILES,
+            description="DramaBox Gemma 3 12B 4-bit encoder",
+        )
+        if not self._is_model_complete(model_dir):
+            raise RuntimeError(f"Downloaded DramaBox model is incomplete: {model_dir}")
+        print(f"✅ DramaBox model ready: {model_dir}")
+        return model_dir
+
+    def _paths_for(self, model_dir: str) -> Dict[str, str]:
+        return {
+            "model_dir": model_dir,
+            "transformer": os.path.join(model_dir, "dramabox-dit-v1.safetensors"),
+            "audio_components": os.path.join(
+                model_dir, "dramabox-audio-components.safetensors"
+            ),
+            "silence_latent": os.path.join(
+                model_dir, "assets", "silence_latent_frame.pt"
+            ),
+            "gemma_root": os.path.join(model_dir, "gemma-3-12b-it-bnb-4bit"),
+        }
+
+    def _is_model_complete(self, model_dir: str) -> bool:
+        if not os.path.isdir(model_dir):
+            return False
+        required = self.DRAMABOX_FILES + [
+            os.path.join("gemma-3-12b-it-bnb-4bit", rel_path)
+            for rel_path in self.GEMMA_FILES
+        ]
+        return all(os.path.isfile(os.path.join(model_dir, path)) for path in required)

engines/dramabox/dramabox_engine.py

@@ -0,0 +1,246 @@
+"""ComfyUI lifecycle wrapper around the official DramaBox warm server."""
+
+import gc
+import importlib.util
+import os
+import sys
+import tempfile
+from typing import Any, Dict, Iterator, Optional
+
+import torch
+import torchaudio
+
+from engines.dramabox.dramabox_downloader import DramaBoxDownloader
+from utils.device import resolve_torch_device
+
+
+class DramaBoxEngine:
+    """Load official DramaBox inference and tear it down cleanly on VRAM clear."""
+
+    SAMPLE_RATE = 48000
+
+    def __init__(
+        self,
+        model_name: str = DramaBoxDownloader.MODEL_NAME,
+        device: str = "auto",
+        precision: str = "auto",
+        model_paths: Optional[Dict[str, str]] = None,
+        memory_mode: str = "fast",
+        transformer_quantization: str = "none",
+        compile_model: bool = False,
+        lora_path: str = "",
+        lora_strength: float = 1.0,
+    ):
+        self.model_name = model_name
+        self.device = resolve_torch_device(device)
+        self.precision = self._resolve_precision(precision)
+        self.model_paths = model_paths
+        self.memory_mode = str(memory_mode)
+        self.transformer_quantization = str(transformer_quantization)
+        self.compile_model = bool(compile_model)
+        self.lora_path = str(lora_path or "").strip()
+        self.lora_strength = float(lora_strength)
+        self._server = None
+        self._server_module = None
+
+    def _resolve_precision(self, precision: str) -> str:
+        value = str(precision or "auto").lower()
+        if value in {"float16", "fp16"}:
+            return "fp16"
+        if value in {"bfloat16", "bf16"}:
+            return "bf16"
+        if torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 8:
+            return "bf16"
+        return "fp16"
+
+    @staticmethod
+    def _vendor_paths():
+        vendor_dir = os.path.join(os.path.dirname(__file__), "vendor")
+        return (
+            os.path.join(vendor_dir, "src"),
+            os.path.join(vendor_dir, "ltx2"),
+        )
+
+    def _import_server(self):
+        if self._server_module is not None:
+            return self._server_module
+
+        src_dir, ltx_dir = self._vendor_paths()
+        for path in (src_dir, ltx_dir):
+            if path not in sys.path:
+                sys.path.insert(0, path)
+
+        module_path = os.path.join(src_dir, "inference_server.py")
+        spec = importlib.util.spec_from_file_location(
+            "tts_audio_suite_dramabox_inference_server", module_path
+        )
+        if spec is None or spec.loader is None:
+            raise ImportError(f"Could not load bundled DramaBox server: {module_path}")
+        module = importlib.util.module_from_spec(spec)
+        sys.modules[spec.name] = module
+        spec.loader.exec_module(module)
+        self._server_module = module
+        return module
+
+    def _ensure_runtime_loaded(self):
+        if self._server is not None:
+            return
+        if not str(self.device).startswith("cuda"):
+            raise RuntimeError(
+                "DramaBox requires an NVIDIA CUDA GPU. Try the experimental "
+                "staged or sequential mode with fp8_cast on lower-memory cards."
+            )
+        if self.model_paths is None:
+            self.model_paths = DramaBoxDownloader().resolve_model_path(self.model_name)
+
+        module = self._import_server()
+        print(
+            f"🔄 Loading DramaBox on {self.device} "
+            f"({self.precision}, official 4-bit Gemma encoder)"
+        )
+        self._server = module.TTSServer(
+            checkpoint=self.model_paths["transformer"],
+            full_checkpoint=self.model_paths["audio_components"],
+            gemma_root=self.model_paths["gemma_root"],
+            device=self.device,
+            dtype=self.precision,
+            compile_model=self.compile_model,
+            bnb_4bit=True,
+            memory_mode=self.memory_mode,
+            transformer_quantization=self.transformer_quantization,
+            lora_path=self.lora_path,
+            lora_strength=self.lora_strength,
+        )
+        print("✅ DramaBox runtime ready")
+
+    @staticmethod
+    def _check_interrupt():
+        import comfy.model_management as model_management
+
+        if model_management.interrupt_processing:
+            raise InterruptedError("DramaBox generation interrupted by user")
+
+    def generate(
+        self,
+        prompt: str,
+        voice_ref_path: Optional[str] = None,
+        cfg_scale: float = 2.5,
+        stg_scale: float = 1.5,
+        duration_multiplier: float = 1.1,
+        gen_duration: float = 0.0,
+        ref_duration: float = 10.0,
+        rescale_scale: Any = "auto",
+        watermark: bool = False,
+        negative_prompt: str = "",
+        seed: int = 42,
+    ) -> Dict[str, Any]:
+        self._ensure_runtime_loaded()
+        self._check_interrupt()
+
+        def progress_callback(_index: int, _total: int, _estimated_seconds: float):
+            self._check_interrupt()
+
+        temp_file = tempfile.NamedTemporaryFile(
+            suffix=".wav", delete=False, prefix="tts_suite_dramabox_"
+        )
+        temp_path = temp_file.name
+        temp_file.close()
+        try:
+            self._server.generate_to_file(
+                prompt=prompt,
+                output=temp_path,
+                voice_ref=voice_ref_path,
+                cfg_scale=float(cfg_scale),
+                stg_scale=float(stg_scale),
+                duration_multiplier=float(duration_multiplier),
+                gen_duration=float(gen_duration),
+                ref_duration=float(ref_duration),
+                rescale_scale=rescale_scale,
+                negative_prompt=str(negative_prompt or ""),
+                seed=int(seed),
+                denoise_ref=False,
+                watermark=bool(watermark),
+                progress_callback=progress_callback,
+            )
+            waveform, sample_rate = torchaudio.load(temp_path)
+        finally:
+            try:
+                os.unlink(temp_path)
+            except OSError:
+                pass
+
+        self._check_interrupt()
+        return {
+            "audio": waveform.detach().float().cpu(),
+            "sample_rate": int(sample_rate),
+        }
+
+    def set_lora(self, lora_path: str = "", strength: float = 1.0, revision: str = ""):
+        """Update the live adapter without rebuilding the base DramaBox runtime."""
+        self.lora_path = str(lora_path or "").strip()
+        self.lora_strength = float(strength)
+        if self._server is not None:
+            self._server.configure_lora(
+                self.lora_path,
+                self.lora_strength,
+                revision=str(revision or ""),
+            )
+
+    def parameters(self) -> Iterator[torch.nn.Parameter]:
+        """Expose loaded submodule parameters for ComfyUI memory accounting."""
+        if self._server is None:
+            return
+        seen = set()
+        stack = list(vars(self._server).values())
+        while stack:
+            value = stack.pop()
+            if id(value) in seen:
+                continue
+            seen.add(id(value))
+            if isinstance(value, torch.nn.Module):
+                yield from value.parameters()
+            elif hasattr(value, "__dict__"):
+                stack.extend(vars(value).values())
+
+    def unload_runtime(self):
+        """Drop the quantized Gemma and LTX runtime instead of copying it to RAM."""
+        server = self._server
+        if server is not None:
+            for name in (
+                "_ref_denoise_cache",
+                "_prompt_encoder",
+                "_velocity_model",
+                "_audio_conditioner",
+                "_audio_decoder",
+                "_ref_denoiser",
+            ):
+                value = getattr(server, name, None)
+                if hasattr(value, "clear"):
+                    try:
+                        value.clear()
+                    except Exception:
+                        pass
+                try:
+                    setattr(server, name, None)
+                except Exception:
+                    pass
+        self._server = None
+        del server
+        gc.collect()
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            if hasattr(torch.cuda, "ipc_collect"):
+                try:
+                    torch.cuda.ipc_collect()
+                except Exception:
+                    pass
+
+    def to(self, device):
+        target = str(device) if isinstance(device, str) else str(torch.device(device))
+        if target.startswith("cpu"):
+            self.unload_runtime()
+        self.device = target
+        return self
+
+    def unload(self):
+        self.unload_runtime()

engines/dramabox/training/__init__.py

@@ -0,0 +1,5 @@
+"""DramaBox LoRA dataset and training integration."""
+
+from .handler import DramaBoxTrainingHandler
+
+__all__ = ["DramaBoxTrainingHandler"]

engines/dramabox/training/dataset.py

@@ -0,0 +1,458 @@
+"""Dataset normalization for the official DramaBox IC-LoRA trainer.
+
+The upstream preprocessor accepts JSONL and TSV, but the upstream training
+loop builds its speaker map from ``~``-delimited index rows.  This module keeps
+that conversion in the suite so a manifest that is valid for preprocessing is
+also valid for training.
+"""
+
+from __future__ import annotations
+
+import csv
+import hashlib
+import json
+import os
+import re
+import wave
+from pathlib import Path
+from typing import Any, Dict, Iterable, List, Optional, Tuple
+
+import folder_paths
+
+
+AUDIO_EXTENSIONS = {".wav", ".flac", ".mp3", ".ogg", ".m4a", ".aac"}
+PREPROCESSED_SAMPLE_PATTERN = re.compile(r"sample_(\d+)\.pt$")
+
+
+def slugify(value: Any) -> str:
+    safe = "".join(
+        ch if ch.isalnum() or ch in ("-", "_") else "_"
+        for ch in str(value or "").strip()
+    )
+    safe = safe.strip("_")
+    return safe or "dramabox_lora"
+
+
+def get_dramabox_training_root() -> str:
+    root = os.path.join(
+        folder_paths.get_output_directory(), "tts_audio_suite_training", "dramabox"
+    )
+    os.makedirs(root, exist_ok=True)
+    return root
+
+
+def _resolve_source_path(value: str) -> Path:
+    raw = os.path.expanduser(str(value or "").strip())
+    if not raw:
+        raise ValueError("dataset_source is required")
+
+    candidates = [Path(raw)]
+    input_root = Path(folder_paths.get_input_directory())
+    candidates.extend((input_root / raw, input_root / "datasets" / raw))
+    for candidate in candidates:
+        if candidate.is_file():
+            return candidate.resolve()
+    raise FileNotFoundError(f"DramaBox dataset source not found: {value}")
+
+
+def _resolve_audio_path(raw_path: Any, *, source_path: Path, audio_dir: str) -> Path:
+    value = os.path.expanduser(str(raw_path or "").strip())
+    if not value:
+        raise ValueError("Dataset row is missing audio_filepath/audio_path")
+
+    candidates: List[Path] = []
+    if os.path.isabs(value):
+        candidates.append(Path(value))
+    else:
+        if audio_dir:
+            candidates.append(Path(os.path.expanduser(audio_dir)) / value)
+        candidates.append(source_path.parent / value)
+        candidates.append(Path(value))
+
+    for candidate in candidates:
+        if candidate.is_file():
+            return candidate.resolve()
+    raise FileNotFoundError(f"DramaBox audio file not found: {raw_path}")
+
+
+def _clean_text(value: Any) -> str:
+    return re.sub(r"\s+", " ", str(value or "").replace("\x00", "")).strip()
+
+
+def _speaker_value(row: Dict[str, Any], default: str = "speaker_1") -> str:
+    value = (
+        row.get("speaker")
+        or row.get("speaker_id")
+        or row.get("voice")
+        or row.get("character")
+        or default
+    )
+    return _clean_text(value).replace("~", "_") or default
+
+
+def _language_value(row: Dict[str, Any]) -> str:
+    return _clean_text(row.get("language") or row.get("lang") or "en").replace("~", "_") or "en"
+
+
+def _coerce_float(value: Any, default: float = 0.0) -> float:
+    try:
+        parsed = float(value)
+    except (TypeError, ValueError):
+        return float(default)
+    return parsed if parsed > 0 else float(default)
+
+
+def _probe_audio(path: Path) -> Tuple[int, int, float]:
+    """Return sample rate, frame count, and duration without loading audio."""
+    try:
+        import torchaudio
+
+        info = torchaudio.info(str(path))
+        sample_rate = int(getattr(info, "sample_rate", 0) or 0)
+        frames = int(getattr(info, "num_frames", 0) or 0)
+        if sample_rate > 0 and frames > 0:
+            return sample_rate, frames, frames / sample_rate
+    except Exception:
+        pass
+
+    if path.suffix.lower() == ".wav":
+        with wave.open(str(path), "rb") as handle:
+            sample_rate = int(handle.getframerate())
+            frames = int(handle.getnframes())
+        if sample_rate > 0 and frames > 0:
+            return sample_rate, frames, frames / sample_rate
+
+    raise RuntimeError(
+        f"Could not inspect audio duration for '{path}'. Add a positive duration "
+        "field to the manifest or install a Torchaudio-compatible decoder."
+    )
+
+
+def _parse_manifest(source_path: Path, audio_dir: str) -> Iterable[Dict[str, Any]]:
+    text = source_path.read_text(encoding="utf-8-sig")
+    stripped = text.lstrip()
+    if stripped.startswith("["):
+        raw_rows = json.loads(text)
+    else:
+        raw_rows = [json.loads(line) for line in text.splitlines() if line.strip()]
+    for row in raw_rows:
+        if not isinstance(row, dict):
+            continue
+        yield {
+            "audio": _resolve_audio_path(
+                row.get("audio_filepath", row.get("audio_path", row.get("audio"))),
+                source_path=source_path,
+                audio_dir=audio_dir,
+            ),
+            "text": _clean_text(row.get("text", row.get("transcript", ""))),
+            "duration": _coerce_float(row.get("duration")),
+            "sample_rate": int(_coerce_float(row.get("sample_rate"))),
+            "samples": int(_coerce_float(row.get("samples", row.get("num_frames")))),
+            "speaker": _speaker_value(row),
+            "language": _language_value(row),
+        }
+
+
+def _parse_tsv(source_path: Path, audio_dir: str) -> Iterable[Dict[str, Any]]:
+    with source_path.open("r", encoding="utf-8-sig", newline="") as handle:
+        for row_number, row in enumerate(csv.reader(handle, delimiter="\t"), start=1):
+            if len(row) < 2:
+                continue
+            yield {
+                "audio": _resolve_audio_path(row[0], source_path=source_path, audio_dir=audio_dir),
+                "text": _clean_text(row[1]),
+                "duration": _coerce_float(row[2]) if len(row) > 2 else 0.0,
+                "sample_rate": 0,
+                "samples": 0,
+                "speaker": _clean_text(row[3]).replace("~", "_") if len(row) > 3 else "speaker_1",
+                "language": _clean_text(row[4]).replace("~", "_") if len(row) > 4 else "en",
+                "row_number": row_number,
+            }
+
+
+def _parse_gemini(source_path: Path, audio_dir: str) -> Iterable[Dict[str, Any]]:
+    for line in source_path.read_text(encoding="utf-8-sig").splitlines():
+        parts = line.strip().split("~")
+        if len(parts) < 8:
+            continue
+        file_id, speaker, language = parts[:3]
+        sample_rate = int(_coerce_float(parts[3], 24000))
+        samples = int(_coerce_float(parts[4]))
+        duration = _coerce_float(parts[5])
+        text = _clean_text(parts[-1])
+        yield {
+            "audio": _resolve_audio_path(file_id, source_path=source_path, audio_dir=audio_dir),
+            "text": text,
+            "duration": duration,
+            "sample_rate": sample_rate,
+            "samples": samples,
+            "speaker": _clean_text(speaker).replace("~", "_") or "speaker_1",
+            "language": _clean_text(language).replace("~", "_") or "en",
+        }
+
+
+def _parse_libriheavy(source_path: Path, audio_dir: str) -> Iterable[Dict[str, Any]]:
+    for line in source_path.read_text(encoding="utf-8-sig").splitlines():
+        parts = line.strip().split("~")
+        if len(parts) < 7:
+            continue
+        file_id, speaker, language = parts[:3]
+        # Format: id~speaker~lang~samples~duration_ms~phonemes~text.
+        sample_rate = 24000
+        samples = int(_coerce_float(parts[3]))
+        duration = _coerce_float(parts[4]) / 1000.0 if len(parts) >= 5 else 0.0
+        yield {
+            "audio": _resolve_audio_path(file_id, source_path=source_path, audio_dir=audio_dir),
+            "text": _clean_text(parts[-1]),
+            "duration": duration,
+            "sample_rate": sample_rate,
+            "samples": samples,
+            "speaker": _clean_text(speaker).replace("~", "_") or "speaker_1",
+            "language": _clean_text(language).replace("~", "_") or "en",
+        }
+
+
+def _raw_rows(source_path: Path, dataset_type: str, audio_dir: str) -> Iterable[Dict[str, Any]]:
+    parsers = {
+        "manifest": _parse_manifest,
+        "tsv": _parse_tsv,
+        "gemini_synthetic": _parse_gemini,
+        "libriheavy": _parse_libriheavy,
+    }
+    try:
+        parser = parsers[str(dataset_type)]
+    except KeyError as exc:
+        raise ValueError(f"Unsupported DramaBox dataset type: {dataset_type}") from exc
+    return parser(source_path, audio_dir)
+
+
+def _fingerprint(source_path: Path, *, dataset_type: str, audio_dir: str, min_duration: float, max_duration: float) -> str:
+    stat = source_path.stat()
+    raw = f"{source_path}|{stat.st_size}|{stat.st_mtime_ns}|{dataset_type}|{audio_dir}|{min_duration}|{max_duration}"
+    return hashlib.sha256(raw.encode("utf-8")).hexdigest()[:16]
+
+
+def _normalize_rows(
+    source_path: Path,
+    *,
+    dataset_type: str,
+    audio_dir: str,
+    min_duration: float,
+    max_duration: float,
+) -> List[Dict[str, Any]]:
+    records: List[Dict[str, Any]] = []
+    for row_index, row in enumerate(_raw_rows(source_path, dataset_type, audio_dir)):
+        text = _clean_text(row.get("text"))
+        if not text:
+            continue
+
+        audio = Path(row["audio"]).resolve()
+        sample_rate = int(row.get("sample_rate") or 0)
+        samples = int(row.get("samples") or 0)
+        duration = _coerce_float(row.get("duration"))
+        if not sample_rate or not samples or not duration:
+            try:
+                probed_rate, probed_samples, probed_duration = _probe_audio(audio)
+                sample_rate = sample_rate or probed_rate
+                samples = samples or probed_samples
+                duration = duration or probed_duration
+            except RuntimeError:
+                if duration <= 0:
+                    raise
+                sample_rate = sample_rate or 24000
+                samples = samples or max(1, round(duration * sample_rate))
+
+        if duration < float(min_duration) or duration > float(max_duration):
+            continue
+        records.append(
+            {
+                "id": f"sample_{row_index:06d}",
+                "audio": str(audio),
+                "text": text,
+                "duration": float(duration),
+                "sample_rate": int(sample_rate),
+                "samples": int(samples),
+                "speaker": _speaker_value(row),
+                "language": _language_value(row),
+            }
+        )
+
+    if not records:
+        raise ValueError(
+            "DramaBox dataset preparation produced no usable rows. Check the audio paths, "
+            "transcripts, and the min/max duration filters."
+        )
+
+    speaker_counts: Dict[str, int] = {}
+    for record in records:
+        speaker_counts[record["speaker"]] = speaker_counts.get(record["speaker"], 0) + 1
+    unusable = sorted(name for name, count in speaker_counts.items() if count < 2)
+    if unusable:
+        raise ValueError(
+            "DramaBox LoRA training needs at least two clips per speaker so the official "
+            f"trainer can choose a reference clip. Speakers with fewer than two clips: {', '.join(unusable)}."
+        )
+    return records
+
+
+def _write_index(records: List[Dict[str, Any]], index_path: Path) -> None:
+    index_path.parent.mkdir(parents=True, exist_ok=True)
+    with index_path.open("w", encoding="utf-8") as handle:
+        for record in records:
+            text = str(record["text"]).replace("\r", " ").replace("\n", " ")
+            handle.write(
+                "~".join(
+                    (
+                        str(Path(record["audio"]).resolve()),
+                        str(record["speaker"]),
+                        str(record["language"]),
+                        str(int(record["sample_rate"])),
+                        str(int(record["samples"])),
+                        f"{float(record['duration']):.6f}",
+                        "_",
+                        text,
+                    )
+                )
+                + "\n"
+            )
+
+
+def _preprocessed_indices(directory: Path) -> set[int]:
+    indices: set[int] = set()
+    if not directory.is_dir():
+        return indices
+    for path in directory.glob("sample_*.pt"):
+        match = PREPROCESSED_SAMPLE_PATTERN.fullmatch(path.name)
+        if match:
+            indices.add(int(match.group(1)))
+    return indices
+
+
+def validate_preprocessed_dataset(
+    records: List[Dict[str, Any]],
+    preprocessed_dir: str | Path,
+    *,
+    raise_on_missing: bool = False,
+) -> bool:
+    """Require matching text conditions and audio latents for every index row."""
+    root = Path(preprocessed_dir)
+    expected = set(range(len(records)))
+    available = _preprocessed_indices(root / "conditions") & _preprocessed_indices(
+        root / "audio_latents"
+    )
+    missing = sorted(expected - available)
+    complete = bool(expected) and not missing
+    if raise_on_missing and not complete:
+        preview = ", ".join(str(index) for index in missing[:10]) or "all"
+        suffix = "..." if len(missing) > 10 else ""
+        raise RuntimeError(
+            "DramaBox preprocessing did not produce matching condition/audio-latent "
+            f"files for {len(missing) or len(expected)} sample(s) (indices: {preview}{suffix}). "
+            "Fix the reported source-audio errors and run Dataset Prep again."
+        )
+    return complete
+
+
+def prepare_dramabox_dataset(
+    shared_settings: Dict[str, Any],
+    *,
+    dataset_source: str,
+    model_name: str,
+    dataset_type: str = "manifest",
+    audio_dir: str = "",
+    min_duration: float = 2.0,
+    max_duration: float = 20.0,
+    reuse_existing: bool = True,
+    preprocess_now: bool = True,
+    dry_run: bool = False,
+) -> Dict[str, Any]:
+    source_path = _resolve_source_path(dataset_source)
+    fingerprint = _fingerprint(
+        source_path,
+        dataset_type=dataset_type,
+        audio_dir=audio_dir,
+        min_duration=min_duration,
+        max_duration=max_duration,
+    )
+    safe_name = slugify(model_name)
+    dataset_root = Path(get_dramabox_training_root()) / "datasets" / f"{safe_name}_{fingerprint}"
+    index_path = dataset_root / "speaker_index.txt"
+    metadata_path = dataset_root / "dataset.json"
+    preprocessed_dir = dataset_root / "preprocessed"
+
+    if reuse_existing and metadata_path.is_file() and index_path.is_file():
+        try:
+            metadata = json.loads(metadata_path.read_text(encoding="utf-8"))
+            records = metadata.get("records") or []
+        except Exception:
+            records = []
+    else:
+        records = []
+
+    if not records:
+        records = _normalize_rows(
+            source_path,
+            dataset_type=dataset_type,
+            audio_dir=audio_dir,
+            min_duration=float(min_duration),
+            max_duration=float(max_duration),
+        )
+        dataset_root.mkdir(parents=True, exist_ok=True)
+        _write_index(records, index_path)
+        metadata_path.write_text(
+            json.dumps(
+                {
+                    "type": "dramabox_dataset",
+                    "source_path": str(source_path),
+                    "dataset_type": dataset_type,
+                    "audio_dir": audio_dir,
+                    "min_duration": float(min_duration),
+                    "max_duration": float(max_duration),
+                    "records": records,
+                },
+                indent=2,
+                ensure_ascii=False,
+            ),
+            encoding="utf-8",
+        )
+
+    # Rewrite cached indexes as well so datasets prepared by older suite
+    # builds migrate from synthetic sample ids to resolvable audio paths.
+    _write_index(records, index_path)
+
+    dataset: Dict[str, Any] = {
+        "type": "training_dataset",
+        "engine_type": "dramabox",
+        "training_mode": "audio_lora",
+        "model_name": model_name,
+        "dataset_type": dataset_type,
+        "source_path": str(source_path),
+        "index_path": str(index_path),
+        "speaker_index": str(index_path),
+        "data_dir": [str(preprocessed_dir)],
+        "preprocessed_dir": str(preprocessed_dir),
+        "min_duration": float(min_duration),
+        "max_duration": float(max_duration),
+        "records": records,
+        "train_records": len(records),
+        "speakers": sorted({str(record["speaker"]) for record in records}),
+        "preprocessed": validate_preprocessed_dataset(records, preprocessed_dir),
+        "dry_run": bool(dry_run),
+        "shared_settings": dict(shared_settings or {}),
+    }
+
+    if preprocess_now and not dry_run and not dataset["preprocessed"]:
+        from .trainer import run_dramabox_preprocess
+
+        run_dramabox_preprocess(dataset, shared_settings, batch_size=8)
+        dataset["preprocessed"] = True
+
+    return dataset
+
+
+__all__ = [
+    "get_dramabox_training_root",
+    "prepare_dramabox_dataset",
+    "slugify",
+    "validate_preprocessed_dataset",
+]

engines/dramabox/training/handler.py

@@ -0,0 +1,82 @@
+"""DramaBox backend for the unified model-training node."""
+
+from __future__ import annotations
+
+from typing import Any, Dict
+
+from engines.training.base_handler import BaseTrainingHandler
+from engines.training.registry import register_training_handler
+
+
+class DramaBoxTrainingHandler(BaseTrainingHandler):
+    engine_type = "dramabox"
+    artifact_type = "lora_adapter"
+
+    def _shared_settings(self, tts_engine: Any) -> Dict[str, Any]:
+        config = self.ensure_engine_type(tts_engine)
+        return {
+            "model_name": config.get("model_name", "DramaBox"),
+            "device": str(config.get("device", "auto")),
+            "precision": str(config.get("precision", "auto")),
+        }
+
+    def build_default_training_config(self, tts_engine: Any) -> Dict[str, Any]:
+        self._shared_settings(tts_engine)
+        return {
+            "type": "training_config",
+            "engine_type": "dramabox",
+            "training_mode": "audio_lora",
+            "base_model": "dev",
+            "steps": 10000,
+            "learning_rate": 1e-4,
+            "lr_scheduler": "cosine",
+            "warmup_steps": 500,
+            "batch_size": 1,
+            "grad_accum": 4,
+            "max_grad_norm": 1.0,
+            "save_every": 500,
+            "log_every": 10,
+            "seed": 42,
+            "lora_rank": 128,
+            "lora_alpha": 128,
+            "lora_dropout": 0.1,
+            "ref_ratio": 0.3,
+            "max_ref_tokens": 200,
+            "text_dropout": 0.4,
+            "preprocess_batch_size": 8,
+            "validation_config": "",
+            "validation_gpu": "",
+            "dry_run": False,
+        }
+
+    def prepare_dataset(self, tts_engine: Any, **kwargs) -> Dict[str, Any]:
+        from .dataset import prepare_dramabox_dataset
+
+        return prepare_dramabox_dataset(self._shared_settings(tts_engine), **kwargs)
+
+    def train(
+        self,
+        tts_engine: Any,
+        training_dataset: Dict[str, Any],
+        training_config: Dict[str, Any],
+        output_name: str = "",
+        resume: bool = False,
+        overwrite: bool = False,
+        continue_from: Any = None,
+        node_id: str = "",
+    ) -> Dict[str, Any]:
+        from .trainer import run_dramabox_training_job
+
+        return run_dramabox_training_job(
+            shared_settings=self._shared_settings(tts_engine),
+            dataset_info=training_dataset,
+            training_config=training_config,
+            output_name=output_name,
+            resume=resume,
+            overwrite=overwrite,
+            continue_from=continue_from,
+            node_id=node_id,
+        )
+
+
+register_training_handler("dramabox", DramaBoxTrainingHandler)

engines/dramabox/training/trainer.py

@@ -0,0 +1,687 @@
+"""Process runner for the official DramaBox IC-LoRA trainer."""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+import shutil
+import subprocess
+import sys
+import time
+from datetime import datetime
+from pathlib import Path
+from typing import Any, Dict, Iterable, Optional
+
+import folder_paths
+
+from engines.dramabox.dramabox_downloader import DramaBoxDownloader
+from engines.training.progress_io import write_json_progress_file
+from engines.training.progress_registry import (
+    finalize_training_job,
+    register_training_job,
+    update_training_job,
+)
+
+from .dataset import (
+    get_dramabox_training_root,
+    slugify,
+    validate_preprocessed_dataset,
+)
+
+
+PROJECT_ROOT = Path(__file__).resolve().parents[3]
+VENDOR_ROOT = PROJECT_ROOT / "engines" / "dramabox" / "vendor"
+PREPROCESS_SCRIPT = VENDOR_ROOT / "src" / "preprocess.py"
+TRAIN_SCRIPT = VENDOR_ROOT / "src" / "train.py"
+
+
+def _write_progress(progress_file: str, *, status: str, phase: str, **updates: Any) -> None:
+    payload: Dict[str, Any] = {}
+    if progress_file and os.path.isfile(progress_file):
+        try:
+            with open(progress_file, "r", encoding="utf-8") as handle:
+                existing = json.load(handle)
+                if isinstance(existing, dict):
+                    payload.update(existing)
+        except Exception:
+            pass
+    payload.update(updates)
+    payload["status"] = status
+    payload["phase"] = phase
+    payload["updated_at"] = datetime.now().isoformat()
+    if progress_file:
+        write_json_progress_file(progress_file, payload, default=str)
+
+
+def _interrupt_requested() -> bool:
+    try:
+        import comfy.model_management as model_management
+    except Exception:
+        return False
+    try:
+        return bool(model_management.processing_interrupted())
+    except Exception:
+        return bool(getattr(model_management, "interrupt_processing", False))
+
+
+def _device_environment(shared_settings: Dict[str, Any]) -> Dict[str, str]:
+    env = os.environ.copy()
+    device = str(shared_settings.get("device", "auto") or "auto").strip().lower()
+    if device.startswith("cpu"):
+        # CPU mode is explicit. This also prevents a CUDA-enabled torch build
+        # from silently taking the user's GPU during preprocessing.
+        env["CUDA_VISIBLE_DEVICES"] = ""
+    elif device.startswith("cuda:"):
+        env["CUDA_VISIBLE_DEVICES"] = device.split(":", 1)[1]
+    return env
+
+
+def _run_process(
+    command: Iterable[str],
+    *,
+    cwd: Path,
+    env: Dict[str, str],
+    phase: str,
+    progress_file: str = "",
+    node_id: str = "",
+    total_steps: int = 0,
+) -> None:
+    command = [str(value) for value in command]
+    print(f"🎓 DramaBox {phase} command: {' '.join(command)}")
+    process = subprocess.Popen(
+        command,
+        cwd=str(cwd),
+        env=env,
+        stdout=subprocess.PIPE,
+        stderr=subprocess.STDOUT,
+        text=True,
+        encoding="utf-8",
+        errors="replace",
+        bufsize=1,
+    )
+    tail: list[str] = []
+    recent_loss_trace: list[Dict[str, Any]] = []
+    best_loss: Optional[float] = None
+    try:
+        assert process.stdout is not None
+        for raw_line in process.stdout:
+            line = raw_line.rstrip()
+            if line:
+                telemetry_match = re.fullmatch(
+                    r"TTS_SUITE_PROGRESS\s+step=(\d+)\s+total=(\d+)", line
+                )
+                if telemetry_match is None:
+                    print(f"[DramaBox {phase}] {line}")
+                    tail.append(line)
+                    del tail[:-30]
+
+                if progress_file:
+                    match = telemetry_match or re.search(
+                        r"(?:Step|step)\s+(\d+)(?:/(\d+))?", line
+                    )
+                    if match:
+                        step = int(match.group(1))
+                        parsed_total = int(match.group(2) or total_steps or 0)
+                        overall_progress = (step / parsed_total) if parsed_total else 0.0
+                        progress_updates: Dict[str, Any] = {
+                            "step": step,
+                            "total_steps": parsed_total,
+                            "overall_progress": overall_progress,
+                            "latest_log": line,
+                        }
+                        loss_match = re.search(
+                            r"\bloss=([-+0-9.eE]+)", line, re.IGNORECASE
+                        )
+                        if loss_match:
+                            loss_value = float(loss_match.group(1))
+                            lr_match = re.search(
+                                r"\blr=([-+0-9.eE]+)", line, re.IGNORECASE
+                            )
+                            learning_rate = (
+                                float(lr_match.group(1)) if lr_match else None
+                            )
+                            recent_loss_trace.append(
+                                {"step": step, "total_loss": loss_value}
+                            )
+                            recent_loss_trace = recent_loss_trace[-120:]
+                            best_loss = (
+                                loss_value
+                                if best_loss is None
+                                else min(best_loss, loss_value)
+                            )
+                            progress_updates.update(
+                                latest_loss=loss_value,
+                                best_gen_loss=best_loss,
+                                recent_loss_trace=recent_loss_trace,
+                                current_metrics={
+                                    "loss_gen_all": loss_value,
+                                    "loss_disc_all": 0.0,
+                                    "loss_mel": 0.0,
+                                    "loss_kl": 0.0,
+                                    "loss_fm": 0.0,
+                                    "learning_rate": learning_rate,
+                                },
+                            )
+                        _write_progress(
+                            progress_file,
+                            status="running",
+                            phase=phase,
+                            **progress_updates,
+                        )
+                        update_training_job(
+                            node_id,
+                            status="running",
+                            phase=phase,
+                            **progress_updates,
+                        )
+                    elif "encoding:" in line.lower():
+                        match = re.search(r"(\d+)\s*/\s*(\d+)", line)
+                        if match:
+                            step = int(match.group(1))
+                            parsed_total = int(match.group(2))
+                            overall_progress = step / max(parsed_total, 1)
+                            _write_progress(
+                                progress_file,
+                                status="running",
+                                phase=phase,
+                                step=step,
+                                total_steps=parsed_total,
+                                overall_progress=overall_progress,
+                                latest_log=line,
+                            )
+                            update_training_job(
+                                node_id,
+                                status="running",
+                                phase=phase,
+                                step=step,
+                                total_steps=parsed_total,
+                                overall_progress=overall_progress,
+                                latest_log=line,
+                            )
+
+            if _interrupt_requested():
+                process.terminate()
+                try:
+                    process.wait(timeout=10)
+                except subprocess.TimeoutExpired:
+                    process.kill()
+                raise InterruptedError(f"DramaBox {phase} interrupted by user")
+
+        return_code = process.wait()
+    except BaseException:
+        if process.poll() is None:
+            process.terminate()
+        raise
+
+    if return_code != 0:
+        details = "\n".join(tail[-10:])
+        raise RuntimeError(
+            f"DramaBox {phase} process failed with exit code {return_code}."
+            + (f"\nLast output:\n{details}" if details else "")
+        )
+
+
+def _resolve_model_paths(shared_settings: Dict[str, Any]) -> Dict[str, str]:
+    model_name = str(shared_settings.get("model_name", "DramaBox") or "DramaBox")
+    return DramaBoxDownloader().resolve_model_path(model_name)
+
+
+def build_preprocess_command(
+    dataset_info: Dict[str, Any],
+    shared_settings: Dict[str, Any],
+    *,
+    batch_size: int = 8,
+    skip_existing: bool = True,
+) -> list[str]:
+    paths = _resolve_model_paths(shared_settings)
+    command = [
+        sys.executable,
+        str(PREPROCESS_SCRIPT),
+        "--dataset-type",
+        "gemini_synthetic",
+        "--index",
+        str(dataset_info["index_path"]),
+        "--output-dir",
+        str(dataset_info["preprocessed_dir"]),
+        "--checkpoint",
+        paths["audio_components"],
+        "--audio-only-ckpt",
+        paths["audio_components"],
+        "--gemma-root",
+        paths["gemma_root"],
+        "--max-duration",
+        str(float(dataset_info.get("max_duration", 20.0))),
+        "--min-duration",
+        str(float(dataset_info.get("min_duration", 2.0))),
+        "--batch-size",
+        str(max(1, int(batch_size))),
+    ]
+    if skip_existing:
+        command.append("--skip-existing")
+    return command
+
+
+def run_dramabox_preprocess(
+    dataset_info: Dict[str, Any],
+    shared_settings: Dict[str, Any],
+    *,
+    batch_size: int = 8,
+    progress_file: str = "",
+    node_id: str = "",
+) -> Dict[str, Any]:
+    command = build_preprocess_command(
+        dataset_info,
+        shared_settings,
+        batch_size=batch_size,
+        skip_existing=True,
+    )
+    _run_process(
+        command,
+        cwd=VENDOR_ROOT,
+        env=_device_environment(shared_settings),
+        phase="preprocess",
+        progress_file=progress_file,
+        node_id=node_id,
+    )
+    validate_preprocessed_dataset(
+        dataset_info.get("records") or [],
+        dataset_info["preprocessed_dir"],
+        raise_on_missing=True,
+    )
+    dataset_info["preprocessed"] = True
+    return dataset_info
+
+
+def _resolve_validation_config(value: str) -> str:
+    raw = os.path.expanduser(str(value or "").strip())
+    if not raw:
+        return ""
+    candidates = [Path(raw)]
+    if not os.path.isabs(raw):
+        candidates.extend(
+            (
+                Path(folder_paths.get_input_directory()) / raw,
+                VENDOR_ROOT / raw,
+            )
+        )
+    for candidate in candidates:
+        if candidate.is_file():
+            return str(candidate.resolve())
+    raise FileNotFoundError(f"DramaBox validation config not found: {value}")
+
+
+def _validation_gpu(training_device: str, requested_gpu: Any) -> str:
+    value = str(requested_gpu or "").strip()
+    if not value:
+        raise ValueError(
+            "DramaBox validation_config requires validation_gpu because official validation "
+            "runs a second full model process. Reserve a GPU different from the training GPU."
+        )
+    if not value.isdigit():
+        raise ValueError("DramaBox validation_gpu must be a non-negative CUDA device index")
+    device = str(training_device or "auto").strip().lower()
+    training_gpu = device.split(":", 1)[1] if device.startswith("cuda:") else "0"
+    if value == training_gpu:
+        raise ValueError(
+            f"DramaBox validation_gpu ({value}) must differ from the training GPU ({training_gpu})"
+        )
+    return value
+
+
+def _resolve_continue_lora(continue_from: Any) -> str:
+    if continue_from is None:
+        return ""
+    if isinstance(continue_from, str):
+        value = os.path.abspath(os.path.expanduser(continue_from.strip()))
+    elif isinstance(continue_from, dict):
+        if str(continue_from.get("engine_type", "") or "").strip().lower() not in {"", "dramabox"}:
+            raise ValueError("continue_from TRAINING_ARTIFACTS must come from a DramaBox training run")
+        value = str(
+            continue_from.get("lora_path")
+            or continue_from.get("model_path")
+            or (continue_from.get("lora_adapter") or {}).get("adapter_path", "")
+        ).strip()
+        value = os.path.abspath(os.path.expanduser(value)) if value else ""
+    else:
+        raise ValueError("Unsupported DramaBox continue_from input")
+
+    if not value:
+        return ""
+    if os.path.isdir(value):
+        candidates = sorted(Path(value).glob("lora_step_*.safetensors"))
+        candidates += [Path(value) / "adapter_model.safetensors"]
+        for candidate in reversed(candidates):
+            if candidate.is_file():
+                return str(candidate)
+        raise FileNotFoundError(f"No DramaBox LoRA weights found in '{value}'")
+    if not os.path.isfile(value):
+        raise FileNotFoundError(f"DramaBox LoRA checkpoint not found: {value}")
+    return value
+
+
+def _managed_lora_root() -> Path:
+    try:
+        from utils.models.extra_paths import get_all_tts_model_paths
+
+        for base_path in get_all_tts_model_paths("TTS"):
+            root = Path(base_path) / "dramabox" / "loras"
+            root.mkdir(parents=True, exist_ok=True)
+            return root
+    except Exception:
+        pass
+    root = Path(folder_paths.models_dir) / "TTS" / "dramabox" / "loras"
+    root.mkdir(parents=True, exist_ok=True)
+    return root
+
+
+def _next_managed_lora_dir(name: str, *, overwrite: bool) -> Path:
+    target = _managed_lora_root() / slugify(name)
+    if overwrite or not target.exists():
+        return target
+    counter = 2
+    while True:
+        candidate = target.parent / f"{target.name}_{counter}"
+        if not candidate.exists():
+            return candidate
+        counter += 1
+
+
+def _latest_lora_file(output_dir: Path) -> Optional[Path]:
+    candidates = sorted(
+        output_dir.glob("lora_step_*.safetensors"),
+        key=lambda path: int(re.search(r"(\d+)", path.stem).group(1))
+        if re.search(r"(\d+)", path.stem)
+        else -1,
+    )
+    if candidates:
+        return candidates[-1]
+    candidate = output_dir / "adapter_model.safetensors"
+    return candidate if candidate.is_file() else None
+
+
+def _build_train_config(
+    dataset_info: Dict[str, Any],
+    shared_settings: Dict[str, Any],
+    training_config: Dict[str, Any],
+    *,
+    output_dir: Path,
+    continue_lora: str,
+    resolve_paths: bool = True,
+) -> Dict[str, Any]:
+    if shared_settings.get("model_paths"):
+        paths = dict(shared_settings["model_paths"])
+    elif resolve_paths:
+        paths = _resolve_model_paths(shared_settings)
+    else:
+        paths = {
+            "transformer": "<dramabox-transformer.safetensors>",
+            "audio_components": "<dramabox-audio-components.safetensors>",
+        }
+    config: Dict[str, Any] = {
+        "data_dir": [str(dataset_info["preprocessed_dir"])],
+        "speaker_index": [str(dataset_info["index_path"])],
+        "output_dir": str(output_dir),
+        "checkpoint": paths["transformer"],
+        "full_checkpoint": paths["audio_components"],
+        "base_model": str(training_config.get("base_model", "dev")),
+        "lora_rank": int(training_config.get("lora_rank", 128)),
+        "lora_alpha": int(training_config.get("lora_alpha", 128)),
+        "lora_dropout": float(training_config.get("lora_dropout", 0.1)),
+        "ref_ratio": float(training_config.get("ref_ratio", 0.3)),
+        "max_ref_tokens": int(training_config.get("max_ref_tokens", 200)),
+        "text_dropout": float(training_config.get("text_dropout", 0.4)),
+        "steps": int(training_config.get("steps", 10000)),
+        "lr": float(training_config.get("learning_rate", 1e-4)),
+        "lr_scheduler": str(training_config.get("lr_scheduler", "cosine")),
+        "warmup_steps": int(training_config.get("warmup_steps", 500)),
+        "batch_size": int(training_config.get("batch_size", 1)),
+        "grad_accum": int(training_config.get("grad_accum", 4)),
+        "max_grad_norm": float(training_config.get("max_grad_norm", 1.0)),
+        "save_every": max(1, int(training_config.get("save_every", 500))),
+        "log_every": int(training_config.get("log_every", 10)),
+        "seed": int(training_config.get("seed", 42)),
+    }
+    if continue_lora:
+        config["resume_lora"] = continue_lora
+    validation_config = _resolve_validation_config(
+        training_config.get("validation_config", "")
+    )
+    if validation_config:
+        config["val_config"] = validation_config
+    return config
+
+
+def _accelerate_command() -> list[str]:
+    executable = shutil.which("accelerate")
+    if executable:
+        return [executable, "launch", "--num_processes", "1"]
+    return [sys.executable, "-m", "accelerate.commands.launch", "--num_processes", "1"]
+
+
+def run_dramabox_training_job(
+    shared_settings: Dict[str, Any],
+    dataset_info: Dict[str, Any],
+    training_config: Dict[str, Any],
+    *,
+    output_name: str = "",
+    resume: bool = False,
+    overwrite: bool = False,
+    continue_from: Any = None,
+    node_id: str = "",
+) -> Dict[str, Any]:
+    if str(dataset_info.get("engine_type", "") or "").strip().lower() != "dramabox":
+        raise ValueError("DramaBox training requires a DramaBox TRAINING_DATASET payload")
+    if str(training_config.get("training_mode", "audio_lora") or "").strip().lower() != "audio_lora":
+        raise ValueError("DramaBox training currently supports audio_lora mode only")
+    if resume:
+        raise RuntimeError(
+            "DramaBox does not support exact optimizer-state resume. Use continue_from with a saved LoRA checkpoint for a warm start."
+        )
+    if str(shared_settings.get("device", "auto") or "auto").strip().lower().startswith("cpu") and not bool(
+        training_config.get("dry_run", False)
+    ):
+        raise RuntimeError(
+            "DramaBox model training requires CUDA. Use dry_run for CPU-only validation; "
+            "no model weights or CUDA process will be started in that mode."
+        )
+    requested_validation = str(
+        training_config.get("validation_config", "") or ""
+    ).strip()
+    if requested_validation:
+        _resolve_validation_config(requested_validation)
+        _validation_gpu(
+            shared_settings.get("device", "auto"),
+            training_config.get("validation_gpu", ""),
+        )
+
+    safe_name = slugify(output_name or dataset_info.get("model_name") or "dramabox_lora")
+    root = Path(get_dramabox_training_root()) / "jobs"
+    root.mkdir(parents=True, exist_ok=True)
+    fingerprint = f"{safe_name}|{dataset_info.get('index_path')}|{training_config}"
+    job_hash = __import__("hashlib").sha256(fingerprint.encode("utf-8")).hexdigest()[:12]
+    job_dir = root / f"{safe_name}_{job_hash}"
+    if job_dir.exists() and not overwrite:
+        job_dir = root / f"{safe_name}_{job_hash}_{int(time.time())}"
+    if overwrite and job_dir.exists():
+        shutil.rmtree(job_dir)
+    job_dir.mkdir(parents=True, exist_ok=True)
+    train_output_dir = job_dir / "lora"
+    progress_file = str(job_dir / "progress.json")
+    managed_dir = _next_managed_lora_dir(safe_name, overwrite=overwrite)
+    continue_lora = _resolve_continue_lora(continue_from)
+
+    register_training_job(
+        node_id,
+        engine_type="dramabox",
+        progress_file=progress_file,
+        job_dir=str(job_dir),
+        model_name=safe_name,
+        sample_rate="48k",
+        total_epochs=1,
+    )
+    try:
+        _write_progress(
+            progress_file,
+            status="starting",
+            phase="setup",
+            engine_type="dramabox",
+            model_name=safe_name,
+            dataset_records=int(dataset_info.get("train_records", 0)),
+            speakers=dataset_info.get("speakers", []),
+            started_at=time.time(),
+        )
+
+        if not bool(dataset_info.get("preprocessed")):
+            if bool(training_config.get("dry_run", False)):
+                print("🧪 DramaBox dry-run: skipping GPU dataset preprocessing")
+            else:
+                _write_progress(progress_file, status="running", phase="preprocess")
+                run_dramabox_preprocess(
+                    dataset_info,
+                    shared_settings,
+                    batch_size=int(training_config.get("preprocess_batch_size", 8)),
+                    progress_file=progress_file,
+                    node_id=node_id,
+                )
+
+        train_config = _build_train_config(
+            dataset_info,
+            shared_settings,
+            training_config,
+            output_dir=train_output_dir,
+            continue_lora=continue_lora,
+            resolve_paths=not bool(training_config.get("dry_run", False)),
+        )
+        config_path = job_dir / "training_config.yaml"
+        import yaml
+
+        config_path.write_text(yaml.safe_dump(train_config, sort_keys=False), encoding="utf-8")
+        (job_dir / "resolved_training_config.json").write_text(
+            json.dumps(
+                {
+                    "dataset": dataset_info,
+                    "shared_settings": shared_settings,
+                    "training_config": training_config,
+                    "official_config": train_config,
+                    "continue_from": continue_lora,
+                },
+                indent=2,
+                ensure_ascii=False,
+                default=str,
+            ),
+            encoding="utf-8",
+        )
+
+        command = [*_accelerate_command(), str(TRAIN_SCRIPT), "--config", str(config_path)]
+        if bool(training_config.get("dry_run", False)):
+            summary = (
+                f"DramaBox dry-run ready: {safe_name} | {dataset_info.get('train_records', 0)} rows | "
+                f"official command prepared without loading CUDA or model weights"
+            )
+            _write_progress(
+                progress_file,
+                status="completed",
+                phase="dry_run",
+                overall_progress=1.0,
+                summary=summary,
+                command=command,
+            )
+            finalize_training_job(node_id, status="completed", summary=summary, dry_run=True)
+            return {
+                "type": "training_artifacts",
+                "engine_type": "dramabox",
+                "training_mode": "audio_lora",
+                "dry_run": True,
+                "job_dir": str(job_dir),
+                "training_config": str(config_path),
+                "summary": summary,
+                "command": command,
+            }
+
+        _write_progress(progress_file, status="running", phase="train", total_steps=int(train_config["steps"]))
+        train_env = _device_environment(shared_settings)
+        if train_config.get("val_config"):
+            paths = _resolve_model_paths(shared_settings)
+            train_env["LTX_CHECKPOINT"] = paths["transformer"]
+            train_env["LTX_FULL_CHECKPOINT"] = paths["audio_components"]
+            train_env["GEMMA_ROOT"] = paths["gemma_root"]
+            train_env["TRAIN_VAL_GPU"] = _validation_gpu(
+                shared_settings.get("device", "auto"),
+                training_config.get("validation_gpu", ""),
+            )
+        _run_process(
+            command,
+            cwd=VENDOR_ROOT,
+            env=train_env,
+            phase="train",
+            progress_file=progress_file,
+            node_id=node_id,
+            total_steps=int(train_config["steps"]),
+        )
+
+        selected_lora = _latest_lora_file(train_output_dir)
+        if selected_lora is None:
+            raise RuntimeError(
+                f"DramaBox training exited successfully but produced no LoRA file in '{train_output_dir}'."
+            )
+        if managed_dir.exists():
+            shutil.rmtree(managed_dir)
+        managed_dir.mkdir(parents=True, exist_ok=True)
+        managed_lora = managed_dir / selected_lora.name
+        shutil.copy2(selected_lora, managed_lora)
+        if selected_lora.name != "adapter_model.safetensors":
+            shutil.copy2(selected_lora, managed_dir / "adapter_model.safetensors")
+        adapter_config = train_output_dir / "adapter_config.json"
+        if adapter_config.is_file():
+            shutil.copy2(adapter_config, managed_dir / adapter_config.name)
+        shutil.copy2(config_path, managed_dir / "training_config.yaml")
+
+        summary = (
+            f"DramaBox audio LoRA training complete: {safe_name} | "
+            f"steps={train_config['steps']} | adapter={managed_lora}"
+        )
+        _write_progress(
+            progress_file,
+            status="completed",
+            phase="done",
+            overall_progress=1.0,
+            output_adapter=str(managed_lora),
+            output_dir=str(managed_dir),
+            summary=summary,
+        )
+        finalize_training_job(
+            node_id,
+            status="completed",
+            output_adapter=str(managed_lora),
+            summary=summary,
+        )
+        return {
+            "type": "training_artifacts",
+            "engine_type": "dramabox",
+            "training_mode": "audio_lora",
+            "model_path": str(managed_dir),
+            "lora_path": str(managed_lora),
+            "job_dir": str(job_dir),
+            "summary": summary,
+            "lora_adapter": {
+                "type": "dramabox_lora",
+                "adapter_path": str(managed_lora),
+                "adapter_dir": str(managed_dir),
+            },
+        }
+    except InterruptedError as error:
+        _write_progress(progress_file, status="cancelled", phase="cancelled", error=str(error))
+        finalize_training_job(node_id, status="cancelled", error=str(error))
+        raise
+    except Exception as error:
+        _write_progress(progress_file, status="error", phase="error", error=str(error))
+        finalize_training_job(node_id, status="error", error=str(error))
+        raise
+
+
+__all__ = [
+    "build_preprocess_command",
+    "run_dramabox_preprocess",
+    "run_dramabox_training_job",
+]

engines/dramabox/vendor/LICENSE

@@ -0,0 +1,381 @@
+                         LTX-2 Community License Agreement
+                           License date: January 5, 2026
+
+
+By using or distributing any portion or element of LTX-2, you agree
+to be bound by this Agreement.
+
+   1. Definitions.
+
+      "Agreement" means the terms and conditions for the license, use,
+      reproduction, and distribution of LTX-2 and the Complementary
+      Materials, as specified in this document.
+
+      "Control" means the direct or indirect ownership of more than
+      fifty percent (50%) of the voting securities or other ownership
+      interests, or the power to direct the management and policies of
+      such Entity through voting rights, contract, or otherwise.
+
+      "Data" means a collection of information and/or content extracted
+      from the dataset used with LTX-2, including to train, pretrain,
+      or otherwise evaluate LTX-2. The Data is not licensed under this
+      Agreement.
+
+      "Derivatives of LTX-2" means all modifications to LTX-2, works
+      based on LTX-2, or any other model which is created or initialized
+      by transfer of patterns of the weights, parameters, activations or
+      output of LTX-2, to the other model, in order to cause the other
+      model to perform similarly to LTX-2, including – but not limited
+      to - distillation methods entailing the use of intermediate data
+      representations or methods based on the generation of synthetic
+      data by LTX-2 for training the other model. For clarity, Derivatives
+      of LTX-2 include: (i) any fine-tuned or adapted weights, parameters,
+      or checkpoints derived from LTX-2; (ii) derivative model architectures
+      that incorporate or are based upon LTX-2's architecture; and
+      (iii) any modified or extended versions of the Complementary
+      Materials. All intellectual property rights in Derivatives of LTX-2
+      shall be subject to the terms of this Agreement, and you may not
+      claim exclusive ownership rights in any Derivatives of LTX-2 that
+      would restrict the rights granted herein.
+
+      "Entity" means any individual, corporation, partnership, limited
+      liability company, or other legal entity. For purposes of this
+      Agreement, an Entity shall be deemed to include, on an aggregative
+      basis, all subsidiaries, affiliates, and other companies under
+      common Control with such Entity. When determining whether an Entity
+      meets any threshold under this Agreement (including revenue
+      thresholds), all subsidiaries, affiliates, and companies under
+      common Control shall be considered collectively.
+
+      "Harm" includes but is not limited to physical, mental,
+      psychological, financial and reputational damage, pain, or loss.
+
+      "Licensor" or "Lightricks" means the owner that is granting the
+      license under this Agreement. For the purposes of this Agreement,
+      the Licensor is Lightricks Ltd.
+
+      "LTX-2" means the large language models, text/image/video/audio/3D
+      generation models, and multimodal large language models and their
+      software and algorithms, including trained model weights, parameters
+      (including optimizer states), machine-learning model code,
+      inference-enabling code, training-enabling code, fine-tuning
+      enabling code, accompanying source code, scripts, documentation,
+      tutorials, examples, and all other elements of the foregoing
+      distributed and made publicly available by Lightricks (including,
+      for example, at https://github.com/Lightricks/LTX-2) for the LTX-2
+      model released on January 5, 2026. This license is applicable to
+      all LTX-2 versions released since January 5, 2026, and all future
+      releases of LTX-2 under this license.
+
+      "Output" means the results of operating LTX-2 as embodied in
+      informational content resulting therefrom.
+
+      "you" (or "your") means an individual or legal Entity licensing
+      LTX-2 in accordance with this Agreement and/or making use of LTX-2
+      for whichever purpose and in any field of use, including usage of
+      LTX-2 in an end-use application - e.g. chatbot, translator, image
+      generator.
+
+   2. Grant of License. Subject to the terms and conditions of this
+      Agreement, you are granted a non-exclusive, worldwide,
+      non-transferable and royalty-free limited license under Licensor's
+      intellectual property or other rights owned by Licensor embodied
+      in LTX-2 to use, reproduce, prepare, distribute, publicly display,
+      publicly perform, sublicense, copy, create derivative works of,
+      and make modifications to LTX-2, for any purpose, subject to the
+      restrictions set forth in Attachment A; provided however, that
+      Entities with annual revenues of at least $10,000,000 (the
+      "Commercial Entities") are required to obtain a paid commercial
+      use license in order to use LTX-2 and Derivatives of LTX-2,
+      subject to the terms and provisions of a different license (the
+      "Commercial Use Agreement"), as will be provided by the Licensor.
+      Commercial Entities interested in such a commercial license are
+      required to [contact Licensor](https://ltx.io/model/licensing).
+      Any commercial use of LTX-2 or Derivatives of LTX-2 by the
+      Commercial Entities not in accordance with this Agreement and/or
+      the Commercial Use Agreement is strictly prohibited and shall be
+      deemed a material breach of this Agreement. Such material breach
+      will be subject, in addition to any license fees owed to Licensor
+      for the period such Commercial Entity used LTX-2 (as will be
+      determined by Licensor), to liquidated damages, which will be paid
+      to Licensor immediately upon demand, in an amount equal to double
+      the amount that would otherwise have been paid by you for the
+      relevant period of time. Such amount reflects a reasonable estimation
+      of the losses and administrative costs incurred due to such breach.
+      You agree and understand that this remedy does not limit the Licensor's
+      right to pursue other remedies available at law or equity.
+
+   3. Distribution and Redistribution. You may host for third parties
+      remote access purposes (e.g. software-as-a-service), reproduce
+      and distribute copies of LTX-2 or Derivatives of LTX-2 thereof in
+      any medium, with or without modifications, provided that you meet
+      the following conditions:
+
+      (a) Use-based restrictions as referenced in paragraph 4 and all
+          provisions of Attachment A MUST be included as an enforceable
+          provision by you in any type of legal agreement (e.g. a
+          license) governing the use and/or distribution of LTX-2 or
+          Derivatives of LTX-2, and you shall give notice to subsequent
+          users you distribute to, that LTX-2 or Derivatives of LTX-2
+          are subject to paragraph 4 and Attachment A in their entirety,
+          including all use restrictions and acceptable use policies;
+
+      (b) You must provide any third party recipients of LTX-2 or
+          Derivatives of LTX-2 a copy of this Agreement, including all
+          attachments and use policies. Any Derivative of LTX-2 (as
+          defined in Section 1, including but not limited to fine-tuned
+          weights, modified training code, models trained on Outputs, or
+          any other derivative) must be distributed exclusively under
+          the terms of this Agreement with a complete copy of this
+          license included;
+
+      (c) You must cause any modified files to carry prominent notices
+          stating that you changed the files;
+
+      (d) You must retain all copyright, patent, trademark, and
+          attribution notices excluding those notices that do not
+          pertain to any part of LTX-2, Derivatives of LTX-2.
+
+      You may add your own copyright statement to your modifications and
+      may provide additional or different license terms and conditions -
+      respecting paragraph 3(a) - for use, reproduction, or distribution
+      of your modifications, or for any such Derivatives of LTX-2 as a
+      whole, provided your use, reproduction, and distribution of LTX-2
+      otherwise complies with the conditions stated in this Agreement,
+      and you provide a complete copy of this Agreement with any such
+      use, reproduction and distribution of LTX-2 and any Derivatives
+      thereof.
+
+   4. Use-based restrictions. The restrictions set forth in Attachment A
+      are considered Use-based restrictions. Therefore, you cannot use
+      LTX-2 and the Derivatives of LTX-2 in violation of the specified
+      restricted uses. You may use LTX-2 subject to this Agreement,
+      including only for lawful purposes and in accordance with the
+      Agreement. "Use" may include creating any content with, fine-tuning,
+      updating, running, training, evaluating and/or re-parametrizing
+      LTX-2. You shall require all of your users who use LTX-2 or a
+      Derivative of LTX-2 to comply with the terms of this paragraph 4.
+
+   5. The Output You Generate. Except as set forth herein, Licensor
+      claims no rights in the Output you generate using LTX-2. You are
+      accountable for input you insert into LTX-2, the Output you
+      generate and its subsequent uses. No use of the Output can
+      contravene any provision as stated in the Agreement.
+
+   6. Updates and Runtime Restrictions. To the maximum extent permitted
+      by law, Licensor reserves the right to restrict (remotely or
+      otherwise) usage of LTX-2 in violation of this Agreement, update
+      LTX-2 through electronic means, or modify the Output of LTX-2
+      based on updates. You shall undertake reasonable efforts to use
+      the latest version of LTX-2. Any use of the non-current version
+      of LTX-2 is done solely at your risk.
+
+   7. Export Controls and Sanctions Compliance. You acknowledge that
+      LTX-2, Derivatives of LTX-2 may be subject to export control laws
+      and regulations, including but not limited to the U.S. Export
+      Administration Regulations and sanctions programs administered by
+      the Office of Foreign Assets Control (OFAC). You represent and
+      warrant that you and any users of LTX-2 are not (i) located in,
+      organized under the laws of, or ordinarily resident in any country
+      or territory subject to comprehensive sanctions; (ii) identified
+      on any U.S. government restricted party list, including the
+      Specially Designated Nationals and Blocked Persons List; or
+      (iii) otherwise prohibited from receiving LTX-2 under applicable
+      law. You shall not export, re-export, or transfer LTX-2, directly
+      or indirectly, in violation of any applicable export control or
+      sanctions laws or regulations. You agree to comply with all
+      applicable trade control laws and shall indemnify and hold
+      Licensor harmless from any claims arising from your failure to
+      comply with such laws.
+
+   8. Trademarks and related. Nothing in this Agreement permits you to
+      make use of Licensor's trademarks, trade names, logos or to
+      otherwise suggest endorsement or misrepresent the relationship
+      between the parties; and any rights not expressly granted herein
+      are reserved by the Licensor.
+
+   9. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides LTX-2 on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or
+      conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS
+      FOR A PARTICULAR PURPOSE. You are solely responsible for
+      determining the appropriateness of using or redistributing LTX-2
+      and Derivatives of LTX-2 and assume any risks associated with
+      your exercise of permissions under this Agreement.
+
+   10. Limitation of Liability. In no event and under no legal theory,
+       whether in tort (including negligence), contract, or otherwise,
+       unless required by applicable law (such as deliberate and grossly
+       negligent acts) or agreed to in writing, shall Licensor be liable
+       to you for damages, including any direct, indirect, special,
+       incidental, or consequential damages of any character arising as
+       a result of this Agreement or out of the use or inability to use
+       LTX-2 (including but not limited to damages for loss of goodwill,
+       work stoppage, computer failure or malfunction, or any and all
+       other commercial damages or losses), even if Licensor has been
+       advised of the possibility of such damages.
+
+   11. Accepting Warranty or Additional Liability. While redistributing
+       LTX-2 and Derivatives of LTX-2, you may, provided you do not
+       violate the terms of this Agreement, choose to offer and charge
+       a fee for, acceptance of support, warranty, indemnity, or other
+       liability obligations. However, in accepting such obligations,
+       you may act only on your own behalf and on your sole
+       responsibility, not on behalf of Licensor, and only if you agree
+       to indemnify, defend, and hold Licensor harmless for any liability
+       incurred by, or claims asserted against Licensor, by reason of
+       your accepting any such warranty or additional liability.
+
+   12. Governing Law. This Agreement and all relations, disputes, claims
+       and other matters arising hereunder (including non-contractual
+       disputes or claims) will be governed exclusively by, and construed
+       exclusively in accordance with, the laws of the State of New York.
+       To the extent permitted by law, choice of laws rules and the
+       United Nations Convention on Contracts for the International Sale
+       of Goods will not apply. For the purposes of adjudicating any
+       action or proceeding to enforce the terms of this Agreement, you
+       hereby irrevocably consent to the exclusive jurisdiction of, and
+       venue in, the federal and state courts located in the County of
+       New York within the State of New York. The prevailing party in
+       any claim or dispute between the parties under this Agreement
+       will be entitled to reimbursement of its reasonable attorneys'
+       fees and costs. You hereby waive the right to a trial by jury,
+       to participate in a class or representative action (including in
+       arbitration), or to combine individual proceedings in court or
+       in arbitration without the consent of all parties.
+
+   13. Term and Termination. This Agreement is effective upon your
+       acceptance and continues until terminated. Licensor may terminate
+       this Agreement immediately upon written notice to you if you
+       breach any provision of this Agreement, including but not limited
+       to violations of the use restrictions in Attachment A or
+       unauthorized commercial use. Upon termination: (a) all rights
+       granted to you under this Agreement will immediately cease;
+       (b) you must immediately cease all use of LTX-2 and Derivatives
+       of LTX-2; (c) you must delete or destroy all copies of LTX-2
+       and Derivatives of LTX-2 in your possession or control; and
+       (d) you must notify any third parties to whom you distributed
+       LTX-2 or Derivatives of LTX-2 of the termination. Sections 8-13,
+       and Section 15 shall survive termination of this Agreement.
+       Termination does not relieve you of any obligations incurred
+       prior to termination, including payment obligations under
+       Section 2. In addition, if You commence a lawsuit or other
+       proceedings (including a cross-claim or counterclaim in a lawsuit)
+       against Licensor or any person or entity alleging that LTX-2 or
+       any Output, or any portion of any of the foregoing, infringe any
+       intellectual property or other right owned or licensable by you,
+       then all licenses granted to you under this Agreement shall
+       terminate as of the date such lawsuit or other proceeding is filed.
+
+   14. Disputes and Arbitration. All disputes arising in connection with
+       this Agreement shall be finally settled by arbitration under the
+       Rules of Arbitration of the International Chamber of Commerce
+       ("ICC Rules"), by one (1) arbitrator appointed in accordance with
+       the ICC Rules. The seat of arbitration shall be New York, NY, USA,
+       and the proceedings shall be conducted in English. The arbitrator
+       shall be empowered to grant any relief that a court could grant.
+       Judgment on the arbitration award may be entered by any court
+       having jurisdiction thereof. Each party waives its right to a
+       trial by jury and to participate in any class or representative
+       action.
+
+   15. If any provision of this Agreement is held to be
+       invalid, illegal
+       or unenforceable, the remaining provisions shall be unaffected
+       thereby and remain valid as if such provision had not been set
+       forth herein.
+
+   END OF TERMS AND CONDITIONS
+
+   ATTACHMENT A: Use Restrictions
+
+      When using the Outputs, LTX-2 and any Derivatives thereof, you
+      will comply with the Acceptable Use Policy. In addition, you
+      agree not to use the Outputs, LTX-2 or its Derivatives in any
+      of the following ways:
+
+      1. In any way that violates any applicable national, federal,
+         state, local or international law or regulation;
+
+      2. For the purpose of exploiting, Harming or attempting to
+         exploit or Harm minors in any way;
+
+      3. To generate or disseminate false information and/or content
+         with the purpose of Harming others;
+
+      4. To generate or disseminate personal identifiable information
+         that can be used to Harm an individual;
+
+      5. To generate or disseminate information and/or content (e.g.
+         images, code, posts, articles), and place the information
+         and/or content in any context (e.g. bot generating tweets)
+         without expressly and intelligibly disclaiming that the
+         information and/or content is machine generated;
+
+      6. To defame, disparage or otherwise harass others;
+
+      7. To impersonate or attempt to impersonate (e.g. deepfakes)
+         others without their consent;
+
+      8. For fully automated decision making that adversely impacts an
+         individual's legal rights or otherwise creates or modifies a
+         binding, enforceable obligation;
+
+      9. For any use intended to or which has the effect of
+         discriminating against or Harming individuals or groups based
+         on online or offline social behavior or known or predicted
+         personal or personality characteristics;
+
+      10. To exploit any of the vulnerabilities of a specific group of
+          persons based on their age, social, physical or mental
+          characteristics, in order to materially distort the behavior
+          of a person pertaining to that group in a manner that causes
+          or is likely to cause that person or another person physical
+          or psychological Harm;
+
+      11. For any use intended to or which has the effect of
+          discriminating against individuals or groups based on legally
+          protected characteristics or categories;
+
+      12. To provide medical advice and medical results interpretation;
+
+      13. To generate or disseminate information for the purpose to be
+          used for administration of justice, law enforcement,
+          immigration or asylum processes, such as predicting an
+          individual will commit fraud/crime commitment (e.g. by text
+          profiling, drawing causal relationships between assertions
+          made in documents, indiscriminate and arbitrarily-targeted use);
+
+      14. To generate and/or disseminate malware (including – but not
+          limited to – ransomware) or any other content to be used for
+          the purpose of harming electronic systems;
+
+      15. To engage in, promote, incite, or facilitate discrimination
+          or other unlawful or harmful conduct in the provision of
+          employment, employment benefits, credit, housing, or other
+          essential goods and services;
+
+      16. To engage in, promote, incite, or facilitate the harassment,
+          abuse, threatening, or bullying of individuals or groups of
+          individuals;
+
+      17. For military, warfare, nuclear industries or applications,
+          weapons development, or any use in connection with activities
+          that may cause death, personal injury, or severe physical or
+          environmental damage;
+
+      18. For commercial use only: To train, improve, or fine-tune any
+          other machine learning model, artificial intelligence system,
+          or competing model, except for Derivatives of LTX-2 as
+          expressly permitted under this Agreement;
+
+      19. To circumvent, disable, or interfere with any technical
+          limitations, safety features, content filters, or use
+          restrictions implemented in LTX-2 by Licensor;
+
+      20. To use LTX-2 or Derivatives of LTX-2 in any product, service,
+          or application that directly competes with Licensor's
+          commercial products or services, or is designed to replace or
+          substitute Licensor's offerings in the market, without
+          obtaining a separate commercial license from Licensor.

engines/dramabox/vendor/UPSTREAM.md

@@ -0,0 +1,41 @@
+# Bundled DramaBox inference and training source
+
+This directory contains the inference-critical source copied unchanged from:
+
+- Repository: `https://github.com/resemble-ai/DramaBox`
+- Commit: `a70a5818e103c1c9fef22409c1e0c707ebf4f8a7`
+- License: LTX-2 Community License Agreement in `LICENSE`
+
+The bundled-code changes are marked inline:
+
+- `ltx2/ltx_pipelines/utils/blocks.py`: local-only Gemma loading prevents
+  Transformers from silently downloading outside TTS Audio Suite's organized
+  ComfyUI model directory; staged modes can defer and release the warm prompt
+  encoder between generation stages.
+- `src/inference_server.py`: ComfyUI cancellation exceptions are allowed to
+  propagate from progress callbacks instead of being swallowed; the official
+  negative-prompt, FP8-cast, compile, and staged-memory controls are exposed to
+  the suite wrapper; suite-managed PEFT LoRA loading is added for trained
+  DramaBox audio adapters.
+- `src/validate.py`: validation accepts the suite's separately organized
+  DramaBox transformer and audio-components checkpoints.
+- `src/preprocess.py`: suite-distributed pre-quantized Gemma checkpoints use
+  the same bitsandbytes-aware prompt-encoder loader as DramaBox inference.
+- `src/train.py`: the batch collator lives at module scope so Windows
+  spawn-based DataLoader workers can serialize it; lightweight per-step
+  telemetry keeps the suite's training dashboard current between normal logs.
+- `ltx2/ltx_core/text_encoders/gemma/encoders/encoder_configurator.py`:
+  supports both wrapped and direct SigLIP vision-tower layouts for the suite's
+  newer Transformers runtime.
+
+The official training entry points are also bundled at this pin:
+
+- `src/preprocess.py`
+- `src/train.py`
+- `src/validate.py`
+- `configs/training_args.example.yaml`
+- `configs/val_config.example.yaml`
+
+The suite invokes these scripts through the unified training backend. Apart
+from the documented compatibility patches, training behavior stays upstream;
+dataset normalization, job lifecycle, and UI wiring remain suite-side.

engines/dramabox/vendor/configs/training_args.example.yaml

@@ -0,0 +1,63 @@
+# DramaBox IC-LoRA training config — values become the defaults for
+# `accelerate launch src/train.py --config configs/training_args.example.yaml`.
+# Any flag explicitly passed on the CLI overrides the YAML.
+
+# ── Data ───────────────────────────────────────────────────────────────────
+# One entry per preprocessed dataset (output dirs from src/preprocess.py).
+data_dir:
+  - /path/to/preprocessed_dataset_a/
+  - /path/to/preprocessed_dataset_b/
+
+# One index file per data_dir entry. Each line follows the format you fed to
+# preprocess.py — see README "Prepare your index file".
+speaker_index:
+  - /path/to/preprocessed_dataset_a/index.txt
+  - /path/to/preprocessed_dataset_b/index.txt
+
+# Output directory for LoRA shards + logs (relative paths resolve against the
+# repo root).
+output_dir: tts_iclora_v1
+
+# ── Base model ─────────────────────────────────────────────────────────────
+# Train your LoRA on top of DramaBox itself (recommended) — the trimmed audio
+# components are enough; no need to ship the raw LTX-2.3 base.
+checkpoint: dramabox-dit-v1.safetensors
+full_checkpoint: dramabox-audio-components.safetensors
+base_model: dev          # 'dev' = ShiftedLogitNormal sampler; 'distilled' = DistilledTimestepSampler
+
+# ── LoRA hyperparams (rank == alpha → scale = 1.0) ─────────────────────────
+lora_rank: 128
+lora_alpha: 128
+lora_dropout: 0.1        # ~0.1 helps regularize on small datasets
+
+# Resume an existing LoRA — step number parsed from the filename
+# (e.g. lora_step_05000.safetensors → starts at step 5000).
+# resume_lora: tts_iclora_v0/lora_step_05000.safetensors
+
+# ── Voice-cloning reference tokens ─────────────────────────────────────────
+ref_ratio: 0.3           # fraction of training samples that get a ref-token tail
+max_ref_tokens: 200      # cap on appended ref tokens after patchification
+
+# CFG training: probability of zeroing the text condition (forces reliance on
+# the voice ref / unconditional path).
+text_dropout: 0.4
+
+# ── Schedule ───────────────────────────────────────────────────────────────
+# Cosine + 1e-4 = from-scratch fine-tune.
+# Constant + 1e-5 = polish on top of an existing LoRA (use with `resume_lora`).
+steps: 10000
+lr: 1.0e-04
+lr_scheduler: cosine
+warmup_steps: 500
+
+batch_size: 1
+grad_accum: 4
+max_grad_norm: 1.0
+
+save_every: 500
+log_every: 50
+seed: 53
+
+# Optional per-save-step validation pass. Generates a sample for every speaker
+# in the val_config so you can A/B listen during training.
+# val_config: configs/val_config.example.yaml

engines/dramabox/vendor/configs/val_config.example.yaml

@@ -0,0 +1,25 @@
+# Validation prompts run by src/validate.py at every --save-every checkpoint.
+# Each entry produces one .wav under <output_dir>/val_step_<N>/<name>.wav.
+#
+# Fields:
+#   name      — short tag used as the output filename
+#   prompt    — full DramaBox-style scene prompt
+#   reference — (optional) absolute path to a 10+ s voice reference clip;
+#               omit for prompt-only generation
+
+speakers:
+  - name: villain_growl
+    prompt: 'A shadowy villain speaks with cold menace, "You have entered my domain, mortal." He chuckles darkly, "Such arrogance will be your undoing."'
+    reference: /path/to/voice_refs/male_villain.wav
+
+  - name: tender_whisper
+    prompt: 'A woman speaks tenderly, "It has been a long day, my love." She whispers, "Close your eyes. I am right here."'
+    reference: /path/to/voice_refs/female_warm.wav
+
+  - name: catgirl_giggle
+    prompt: 'A playful girl already mid-giggle, "Hehehe, oh my gosh you should see your face!" She gasps, "Oh my, hehe, I cannot stop!"'
+    # No `reference:` here — pure prompt-driven generation.
+
+  - name: announcer_smug
+    prompt: 'A confident announcer speaks proudly, "And now, the moment you have all been waiting for." He chuckles knowingly, "Heheh."'
+    reference: /path/to/voice_refs/male_announcer.wav

engines/dramabox/vendor/ltx2/ltx_core/batch_split.py

@@ -0,0 +1,95 @@
+"""Batch-splitting adapter for the transformer.
+Wraps an ``X0Model`` (or ``LayerStreamingWrapper``) and splits batched inputs
+into smaller chunks before forwarding, then concatenates the results. This
+controls peak activation memory at the cost of more forward passes.
+The adapter is transparent — it has the same ``forward`` signature as
+``X0Model`` and proxies attribute access to the wrapped model.
+Example
+-------
+>>> from ltx_core.batch_split import BatchSplitAdapter
+>>> adapter = BatchSplitAdapter(model, max_batch_size=1)
+>>> # Receives B=4, runs 4xB=1 internally, returns B=4
+>>> denoised_video, denoised_audio = adapter(video=v_b4, audio=a_b4, perturbations=ptb)
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+import torch
+from torch import nn
+
+from ltx_core.guidance.perturbations import BatchedPerturbationConfig
+from ltx_core.model.transformer.modality import Modality
+
+
+def _split_perturbations(config: BatchedPerturbationConfig, sizes: list[int]) -> list[BatchedPerturbationConfig]:
+    """Split a ``BatchedPerturbationConfig`` along the batch dimension."""
+    it = iter(config.perturbations)
+    return [BatchedPerturbationConfig([next(it) for _ in range(s)]) for s in sizes]
+
+
+def _merge_tensors(tensors: list[torch.Tensor | None]) -> torch.Tensor | None:
+    """Concatenate tensors along batch dim, or return None if all are None."""
+    non_none = [t for t in tensors if t is not None]
+    if not non_none:
+        return None
+    return torch.cat(non_none, dim=0)
+
+
+class BatchSplitAdapter(nn.Module):
+    """Wraps a model and splits batched forward calls into smaller chunks.
+    Has the same ``forward`` signature as ``X0Model``:
+    ``(video, audio, perturbations) -> (denoised_video, denoised_audio)``.
+    Args:
+        model: The model to wrap (``X0Model``, ``LayerStreamingWrapper``, etc.).
+        max_batch_size: Maximum batch size per forward pass. Input batches
+            larger than this are split into sequential chunks.
+    """
+
+    def __init__(self, model: nn.Module, max_batch_size: int) -> None:
+        if max_batch_size < 1:
+            raise ValueError(f"max_batch_size must be >= 1, got {max_batch_size}")
+        super().__init__()
+        self._model = model
+        self._max_batch_size = max_batch_size
+
+    def _get_chunk_sizes(self, batch_size: int) -> list[int]:
+        full, remainder = divmod(batch_size, self._max_batch_size)
+        sizes = [self._max_batch_size] * full
+        if remainder:
+            sizes.append(remainder)
+        return sizes
+
+    def forward(
+        self,
+        video: Modality | None,
+        audio: Modality | None,
+        perturbations: BatchedPerturbationConfig,
+    ) -> tuple[torch.Tensor | None, torch.Tensor | None]:
+        batch_size = (video or audio).latent.shape[0]
+
+        if batch_size <= self._max_batch_size:
+            return self._model(video=video, audio=audio, perturbations=perturbations)
+
+        sizes = self._get_chunk_sizes(batch_size)
+        n = len(sizes)
+
+        v_chunks = video.split(sizes) if video is not None else [None] * n
+        a_chunks = audio.split(sizes) if audio is not None else [None] * n
+        p_chunks = _split_perturbations(perturbations, sizes)
+
+        chunk_results = [
+            self._model(video=vc, audio=ac, perturbations=pc)
+            for vc, ac, pc in zip(v_chunks, a_chunks, p_chunks, strict=True)
+        ]
+
+        results_v, results_a = zip(*chunk_results, strict=True)
+        return _merge_tensors(list(results_v)), _merge_tensors(list(results_a))
+
+    def __getattr__(self, name: str) -> Any:  # noqa: ANN401
+        """Proxy attribute access to the wrapped model."""
+        try:
+            return super().__getattr__(name)
+        except AttributeError:
+            return getattr(self._model, name)

engines/dramabox/vendor/ltx2/ltx_core/components/__init__.py

@@ -0,0 +1,10 @@
+"""
+Diffusion pipeline components.
+Submodules:
+    diffusion_steps - Diffusion stepping algorithms (EulerDiffusionStep)
+    guiders         - Guidance strategies (CFGGuider, STGGuider, APG variants)
+    noisers         - Noise samplers (GaussianNoiser)
+    patchifiers     - Latent patchification (VideoLatentPatchifier, AudioPatchifier)
+    protocols       - Protocol definitions (Patchifier, etc.)
+    schedulers      - Sigma schedulers (LTX2Scheduler, LinearQuadraticScheduler)
+"""

engines/dramabox/vendor/ltx2/ltx_core/components/diffusion_steps.py

@@ -0,0 +1,106 @@
+import torch
+
+from ltx_core.components.protocols import DiffusionStepProtocol
+from ltx_core.utils import to_velocity
+
+
+class EulerDiffusionStep(DiffusionStepProtocol):
+    """
+    First-order Euler method for diffusion sampling.
+    Takes a single step from the current noise level (sigma) to the next by
+    computing velocity from the denoised prediction and applying: sample + velocity * dt.
+    """
+
+    def step(
+        self, sample: torch.Tensor, denoised_sample: torch.Tensor, sigmas: torch.Tensor, step_index: int, **_kwargs
+    ) -> torch.Tensor:
+        sigma = sigmas[step_index]
+        sigma_next = sigmas[step_index + 1]
+        dt = sigma_next - sigma
+        velocity = to_velocity(sample, sigma, denoised_sample)
+
+        return (sample.to(torch.float32) + velocity.to(torch.float32) * dt).to(sample.dtype)
+
+
+class Res2sDiffusionStep(DiffusionStepProtocol):
+    """
+    Second-order diffusion step for res_2s sampling with SDE noise injection.
+    Used by the res_2s denoising loop. Advances the sample from the current
+    sigma to the next by mixing a deterministic update (from the denoised
+    prediction) with injected noise via ``get_sde_coeff``, producing
+    variance-preserving transitions.
+    """
+
+    @staticmethod
+    def get_sde_coeff(
+        sigma_next: torch.Tensor,
+        sigma_up: torch.Tensor | None = None,
+        sigma_down: torch.Tensor | None = None,
+        sigma_max: torch.Tensor | None = None,
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Compute SDE coefficients (alpha_ratio, sigma_down, sigma_up) for the step.
+        Given either ``sigma_down`` or ``sigma_up``, returns the mixing
+        coefficients used for variance-preserving noise injection. If
+        ``sigma_up`` is provided, ``sigma_down`` and ``alpha_ratio`` are
+        derived; if ``sigma_down`` is provided, ``sigma_up`` and
+        ``alpha_ratio`` are derived.
+        """
+        if sigma_down is not None:
+            alpha_ratio = (1 - sigma_next) / (1 - sigma_down)
+            sigma_up = (sigma_next**2 - sigma_down**2 * alpha_ratio**2).clamp(min=0) ** 0.5
+        elif sigma_up is not None:
+            # Fallback to avoid sqrt(neg_num)
+            sigma_up.clamp_(max=sigma_next * 0.9999)
+            sigmax = sigma_max if sigma_max is not None else torch.ones_like(sigma_next)
+            sigma_signal = sigmax - sigma_next
+            sigma_residual = (sigma_next**2 - sigma_up**2).clamp(min=0) ** 0.5
+            alpha_ratio = sigma_signal + sigma_residual
+            sigma_down = sigma_residual / alpha_ratio
+        else:
+            alpha_ratio = torch.ones_like(sigma_next)
+            sigma_down = sigma_next
+            sigma_up = torch.zeros_like(sigma_next)
+
+        sigma_up = torch.nan_to_num(sigma_up if sigma_up is not None else torch.zeros_like(sigma_next), 0.0)
+        # Replace NaNs in sigma_down with corresponding sigma_next elements (float32)
+        nan_mask = torch.isnan(sigma_down)
+        sigma_down[nan_mask] = sigma_next[nan_mask].to(sigma_down.dtype)
+        alpha_ratio = torch.nan_to_num(alpha_ratio, 1.0)
+
+        return alpha_ratio, sigma_down, sigma_up
+
+    def step(
+        self,
+        sample: torch.Tensor,
+        denoised_sample: torch.Tensor,
+        sigmas: torch.Tensor,
+        step_index: int,
+        noise: torch.Tensor,
+        eta: float = 0.5,
+    ) -> torch.Tensor:
+        """Advance one step with SDE noise injection via get_sde_coeff.
+        Args:
+            sample: Current noisy sample.
+            denoised_sample: Denoised prediction from the model.
+            sigmas: Noise schedule tensor.
+            step_index: Current step index in the schedule.
+            noise: Random noise tensor for stochastic injection.
+            eta: Controls stochastic noise injection strength (0=deterministic, 1=maximum). Default 0.5.
+        Returns:
+            Next sample with SDE noise injection applied.
+        """
+        sigma = sigmas[step_index]
+        sigma_next = sigmas[step_index + 1]
+        alpha_ratio, sigma_down, sigma_up = self.get_sde_coeff(sigma_next, sigma_up=sigma_next * eta)
+        output_dtype = denoised_sample.dtype
+        if torch.any(sigma_up == 0) or torch.any(sigma_next == 0):
+            return denoised_sample
+
+        # Extract epsilon prediction
+        eps_next = (sample - denoised_sample) / (sigma - sigma_next)
+        denoised_next = sample - sigma * eps_next
+
+        # Mix deterministic and stochastic components
+        x_noised = alpha_ratio * (denoised_next + sigma_down * eps_next) + sigma_up * noise
+        return x_noised.to(output_dtype)

engines/dramabox/vendor/ltx2/ltx_core/components/guiders.py

@@ -0,0 +1,383 @@
+import math
+from collections.abc import Mapping, Sequence
+from dataclasses import dataclass, field
+
+import torch
+
+from ltx_core.components.protocols import GuiderProtocol
+
+
+@dataclass(frozen=True)
+class CFGGuider(GuiderProtocol):
+    """
+    Classifier-free guidance (CFG) guider.
+    Computes the guidance delta as (scale - 1) * (cond - uncond), steering the
+    denoising process toward the conditioned prediction.
+    Attributes:
+        scale: Guidance strength. 1.0 means no guidance, higher values increase
+            adherence to the conditioning.
+    """
+
+    scale: float
+
+    def delta(self, cond: torch.Tensor, uncond: torch.Tensor) -> torch.Tensor:
+        return (self.scale - 1) * (cond - uncond)
+
+    def enabled(self) -> bool:
+        return self.scale != 1.0
+
+
+@dataclass(frozen=True)
+class CFGStarRescalingGuider(GuiderProtocol):
+    """
+    Calculates the CFG delta between conditioned and unconditioned samples.
+    To minimize offset in the denoising direction and move mostly along the
+    conditioning axis within the distribution, the unconditioned sample is
+    rescaled in accordance with the norm of the conditioned sample.
+    Attributes:
+        scale (float):
+            Global guidance strength. A value of 1.0 corresponds to no extra
+            guidance beyond the base model prediction. Values > 1.0 increase
+            the influence of the conditioned sample relative to the
+            unconditioned one.
+    """
+
+    scale: float
+
+    def delta(self, cond: torch.Tensor, uncond: torch.Tensor) -> torch.Tensor:
+        rescaled_neg = projection_coef(cond, uncond) * uncond
+        return (self.scale - 1) * (cond - rescaled_neg)
+
+    def enabled(self) -> bool:
+        return self.scale != 1.0
+
+
+@dataclass(frozen=True)
+class STGGuider(GuiderProtocol):
+    """
+    Calculates the STG delta between conditioned and perturbed denoised samples.
+    Perturbed samples are the result of the denoising process with perturbations,
+    e.g. attentions acting as passthrough for certain layers and modalities.
+    Attributes:
+        scale (float):
+            Global strength of the STG guidance. A value of 0.0 disables the
+            guidance. Larger values increase the correction applied in the
+            direction of (pos_denoised - perturbed_denoised).
+    """
+
+    scale: float
+
+    def delta(self, pos_denoised: torch.Tensor, perturbed_denoised: torch.Tensor) -> torch.Tensor:
+        return self.scale * (pos_denoised - perturbed_denoised)
+
+    def enabled(self) -> bool:
+        return self.scale != 0.0
+
+
+@dataclass(frozen=True)
+class LtxAPGGuider(GuiderProtocol):
+    """
+    Calculates the APG (adaptive projected guidance) delta between conditioned
+    and unconditioned samples.
+    To minimize offset in the denoising direction and move mostly along the
+    conditioning axis within the distribution, the (cond - uncond) delta is
+    decomposed into components parallel and orthogonal to the conditioned
+    sample. The `eta` parameter weights the parallel component, while `scale`
+    is applied to the orthogonal component. Optionally, a norm threshold can
+    be used to suppress guidance when the magnitude of the correction is small.
+    Attributes:
+        scale (float):
+            Strength applied to the component of the guidance that is orthogonal
+            to the conditioned sample. Controls how aggressively we move in
+            directions that change semantics but stay consistent with the
+            conditioning manifold.
+        eta (float):
+            Weight of the component of the guidance that is parallel to the
+            conditioned sample. A value of 1.0 keeps the full parallel
+            component; values in [0, 1] attenuate it, and values > 1.0 amplify
+            motion along the conditioning direction.
+        norm_threshold (float):
+            Minimum L2 norm of the guidance delta below which the guidance
+            can be reduced or ignored (depending on implementation).
+            This is useful for avoiding noisy or unstable updates when the
+            guidance signal is very small.
+    """
+
+    scale: float
+    eta: float = 1.0
+    norm_threshold: float = 0.0
+
+    def delta(self, cond: torch.Tensor, uncond: torch.Tensor) -> torch.Tensor:
+        guidance = cond - uncond
+        if self.norm_threshold > 0:
+            ones = torch.ones_like(guidance)
+            guidance_norm = guidance.norm(p=2, dim=[-1, -2, -3], keepdim=True)
+            scale_factor = torch.minimum(ones, self.norm_threshold / guidance_norm)
+            guidance = guidance * scale_factor
+        proj_coeff = projection_coef(guidance, cond)
+        g_parallel = proj_coeff * cond
+        g_orth = guidance - g_parallel
+        g_apg = g_parallel * self.eta + g_orth
+
+        return g_apg * (self.scale - 1)
+
+    def enabled(self) -> bool:
+        return self.scale != 1.0
+
+
+@dataclass(frozen=False)
+class LegacyStatefulAPGGuider(GuiderProtocol):
+    """
+    Calculates the APG (adaptive projected guidance) delta between conditioned
+    and unconditioned samples.
+    To minimize offset in the denoising direction and move mostly along the
+    conditioning axis within the distribution, the (cond - uncond) delta is
+    decomposed into components parallel and orthogonal to the conditioned
+    sample. The `eta` parameter weights the parallel component, while `scale`
+    is applied to the orthogonal component. Optionally, a norm threshold can
+    be used to suppress guidance when the magnitude of the correction is small.
+    Attributes:
+        scale (float):
+            Strength applied to the component of the guidance that is orthogonal
+            to the conditioned sample. Controls how aggressively we move in
+            directions that change semantics but stay consistent with the
+            conditioning manifold.
+        eta (float):
+            Weight of the component of the guidance that is parallel to the
+            conditioned sample. A value of 1.0 keeps the full parallel
+            component; values in [0, 1] attenuate it, and values > 1.0 amplify
+            motion along the conditioning direction.
+        norm_threshold (float):
+            Minimum L2 norm of the guidance delta below which the guidance
+            can be reduced or ignored (depending on implementation).
+            This is useful for avoiding noisy or unstable updates when the
+            guidance signal is very small.
+        momentum (float):
+            Exponential moving-average coefficient for accumulating guidance
+            over time. running_avg = momentum * running_avg + guidance
+    """
+
+    scale: float
+    eta: float
+    norm_threshold: float = 5.0
+    momentum: float = 0.0
+    # it is user's responsibility not to use same APGGuider for several denoisings or different modalities
+    # in order not to share accumulated average across different denoisings or modalities
+    running_avg: torch.Tensor | None = None
+
+    def delta(self, cond: torch.Tensor, uncond: torch.Tensor) -> torch.Tensor:
+        guidance = cond - uncond
+        if self.momentum != 0:
+            if self.running_avg is None:
+                self.running_avg = guidance.clone()
+            else:
+                self.running_avg = self.momentum * self.running_avg + guidance
+            guidance = self.running_avg
+
+        if self.norm_threshold > 0:
+            ones = torch.ones_like(guidance)
+            guidance_norm = guidance.norm(p=2, dim=[-1, -2, -3], keepdim=True)
+            scale_factor = torch.minimum(ones, self.norm_threshold / guidance_norm)
+            guidance = guidance * scale_factor
+
+        proj_coeff = projection_coef(guidance, cond)
+        g_parallel = proj_coeff * cond
+        g_orth = guidance - g_parallel
+        g_apg = g_parallel * self.eta + g_orth
+
+        return g_apg * self.scale
+
+    def enabled(self) -> bool:
+        return self.scale != 0.0
+
+
+@dataclass(frozen=True)
+class MultiModalGuiderParams:
+    """
+    Parameters for the multi-modal guider.
+    """
+
+    cfg_scale: float = 1.0
+    "CFG (Classifier-free guidance) scale controlling how strongly the model adheres to the prompt."
+    stg_scale: float = 0.0
+    "STG (Spatio-Temporal Guidance) scale controls how strongly the model reacts to the perturbation of the modality."
+    stg_blocks: list[int] | None = field(default_factory=list)
+    "Which transformer blocks to perturb for STG."
+    rescale_scale: float = 0.0
+    "Rescale scale controlling how strongly the model rescales the modality after applying other guidance."
+    modality_scale: float = 1.0
+    "Modality scale controlling how strongly the model reacts to the perturbation of the modality."
+    cfg_clamp_scale: float = 0.0
+    "Clamp guided prediction std to this multiple of conditioned prediction std. 0 = disabled."
+    skip_step: int = 0
+    "Skip step controlling how often the model skips the step."
+
+
+def _params_for_sigma_from_sorted_dict(
+    sigma: float, params_by_sigma: Sequence[tuple[float, MultiModalGuiderParams]]
+) -> MultiModalGuiderParams:
+    """
+    Return params for the given sigma from a sorted (sigma_upper_bound -> params) structure.
+    Keys are sorted descending (bin upper bounds). Bin i is (key_{i+1}, key_i].
+    Get all keys >= sigma; use last in list (smallest such key = upper bound of bin containing sigma),
+    or last entry in the sequence if list is empty (sigma above max key).
+    """
+    if not params_by_sigma:
+        raise ValueError("params_by_sigma must be non-empty")
+    sigma = float(sigma)
+    keys_desc = [k for k, _ in params_by_sigma]
+    keys_ge_sigma = [k for k in keys_desc if k >= sigma]
+    # sigma above all keys: use first bin (max key)
+    key = keys_ge_sigma[-1] if keys_ge_sigma else keys_desc[0]
+    return next(p for k, p in params_by_sigma if k == key)
+
+
+@dataclass(frozen=True)
+class MultiModalGuider:
+    """
+    Multi-modal guider with constant params per instance.
+    For sigma-dependent params, use MultiModalGuiderFactory.build_from_sigma(sigma) to
+    obtain a guider for each step.
+    """
+
+    params: MultiModalGuiderParams
+    negative_context: torch.Tensor | None = None
+
+    def calculate(
+        self,
+        cond: torch.Tensor,
+        uncond_text: torch.Tensor | float,
+        uncond_perturbed: torch.Tensor | float,
+        uncond_modality: torch.Tensor | float,
+    ) -> torch.Tensor:
+        """
+        The guider calculates the guidance delta as (scale - 1) * (cond - uncond) for cfg and modality cfg,
+        and as scale * (cond - uncond) for stg, steering the denoising process away from the unconditioned
+        prediction.
+        """
+        pred = (
+            cond
+            + (self.params.cfg_scale - 1) * (cond - uncond_text)
+            + self.params.stg_scale * (cond - uncond_perturbed)
+            + (self.params.modality_scale - 1) * (cond - uncond_modality)
+        )
+
+        if self.params.rescale_scale != 0:
+            factor = cond.std() / pred.std()
+            factor = self.params.rescale_scale * factor + (1 - self.params.rescale_scale)
+            pred = pred * factor
+
+        # Clamp guided prediction to prevent trajectory overshoot.
+        # Instead of global std (which averages over all tokens), clamp per-token.
+        # This catches individual tokens that overshoot even if the global std looks fine.
+        if self.params.cfg_clamp_scale > 0:
+            cfg_delta = pred - cond
+            # Per-token magnitude clamping
+            delta_norm = cfg_delta.norm(dim=-1, keepdim=True)  # [B, T, 1]
+            cond_norm = cond.norm(dim=-1, keepdim=True)
+            max_norm = cond_norm * self.params.cfg_clamp_scale
+            # Clamp tokens where delta exceeds max
+            scale = torch.where(
+                delta_norm > max_norm,
+                max_norm / delta_norm.clamp(min=1e-8),
+                torch.ones_like(delta_norm),
+            )
+            pred = cond + cfg_delta * scale
+
+        return pred
+
+    def do_unconditional_generation(self) -> bool:
+        """Returns True if the guider is doing unconditional generation."""
+        return not math.isclose(self.params.cfg_scale, 1.0)
+
+    def do_perturbed_generation(self) -> bool:
+        """Returns True if the guider is doing perturbed generation."""
+        return not math.isclose(self.params.stg_scale, 0.0)
+
+    def do_isolated_modality_generation(self) -> bool:
+        """Returns True if the guider is doing isolated modality generation."""
+        return not math.isclose(self.params.modality_scale, 1.0)
+
+    def should_skip_step(self, step: int) -> bool:
+        """Returns True if the guider should skip the step."""
+        if self.params.skip_step == 0:
+            return False
+        return step % (self.params.skip_step + 1) != 0
+
+
+@dataclass(frozen=True)
+class MultiModalGuiderFactory:
+    """
+    Factory that creates a MultiModalGuider for a given sigma.
+    Single source of truth: _params_by_sigma (schedule). Use constant() for
+    one params for all sigma, from_dict() for sigma-binned params.
+    """
+
+    negative_context: torch.Tensor | None = None
+    _params_by_sigma: tuple[tuple[float, MultiModalGuiderParams], ...] = ()
+
+    @classmethod
+    def constant(
+        cls,
+        params: MultiModalGuiderParams,
+        negative_context: torch.Tensor | None = None,
+    ) -> "MultiModalGuiderFactory":
+        """Build a factory with constant params (same guider for all sigma)."""
+        return cls(
+            negative_context=negative_context,
+            _params_by_sigma=((float("inf"), params),),
+        )
+
+    @classmethod
+    def from_dict(
+        cls,
+        sigma_to_params: Mapping[float, MultiModalGuiderParams],
+        negative_context: torch.Tensor | None = None,
+    ) -> "MultiModalGuiderFactory":
+        """
+        Build a factory from a dict of sigma_value -> MultiModalGuiderParams.
+        Keys are sorted descending and used for bin lookup in params(sigma).
+        """
+        if not sigma_to_params:
+            raise ValueError("sigma_to_params must be non-empty")
+        sorted_items = tuple(sorted(sigma_to_params.items(), key=lambda x: x[0], reverse=True))
+        return cls(negative_context=negative_context, _params_by_sigma=sorted_items)
+
+    def params(self, sigma: float | torch.Tensor) -> MultiModalGuiderParams:
+        """Return params effective for the given sigma (getter; single source of truth)."""
+        sigma_val = float(sigma.item() if isinstance(sigma, torch.Tensor) else sigma)
+        return _params_for_sigma_from_sorted_dict(sigma_val, self._params_by_sigma)
+
+    def build_from_sigma(self, sigma: float | torch.Tensor) -> MultiModalGuider:
+        """Return a MultiModalGuider with params effective for the given sigma."""
+        return MultiModalGuider(
+            params=self.params(sigma),
+            negative_context=self.negative_context,
+        )
+
+
+def create_multimodal_guider_factory(
+    params: MultiModalGuiderParams | MultiModalGuiderFactory,
+    negative_context: torch.Tensor | None = None,
+) -> MultiModalGuiderFactory:
+    """
+    Create or return a MultiModalGuiderFactory. Pass constant params for a
+    single-params factory (uses MultiModalGuiderFactory.constant), or an existing
+    MultiModalGuiderFactory. When given a factory, returns it as-is unless
+    negative_context is provided. For sigma-dependent params use
+    MultiModalGuiderFactory.from_dict(...) and pass that as params.
+    """
+    if isinstance(params, MultiModalGuiderFactory):
+        if negative_context is not None and params.negative_context is not negative_context:
+            return MultiModalGuiderFactory.from_dict(dict(params._params_by_sigma), negative_context=negative_context)
+        return params
+    return MultiModalGuiderFactory.constant(params, negative_context=negative_context)
+
+
+def projection_coef(to_project: torch.Tensor, project_onto: torch.Tensor) -> torch.Tensor:
+    batch_size = to_project.shape[0]
+    positive_flat = to_project.reshape(batch_size, -1)
+    negative_flat = project_onto.reshape(batch_size, -1)
+    dot_product = torch.sum(positive_flat * negative_flat, dim=1, keepdim=True)
+    squared_norm = torch.sum(negative_flat**2, dim=1, keepdim=True) + 1e-8
+    return dot_product / squared_norm

engines/dramabox/vendor/ltx2/ltx_core/components/noisers.py

@@ -0,0 +1,35 @@
+from dataclasses import replace
+from typing import Protocol
+
+import torch
+
+from ltx_core.types import LatentState
+
+
+class Noiser(Protocol):
+    """Protocol for adding noise to a latent state during diffusion."""
+
+    def __call__(self, latent_state: LatentState, noise_scale: float) -> LatentState: ...
+
+
+class GaussianNoiser(Noiser):
+    """Adds Gaussian noise to a latent state, scaled by the denoise mask."""
+
+    def __init__(self, generator: torch.Generator):
+        super().__init__()
+
+        self.generator = generator
+
+    def __call__(self, latent_state: LatentState, noise_scale: float = 1.0) -> LatentState:
+        noise = torch.randn(
+            *latent_state.latent.shape,
+            device=latent_state.latent.device,
+            dtype=latent_state.latent.dtype,
+            generator=self.generator,
+        )
+        scaled_mask = latent_state.denoise_mask * noise_scale
+        latent = noise * scaled_mask + latent_state.latent * (1 - scaled_mask)
+        return replace(
+            latent_state,
+            latent=latent.to(latent_state.latent.dtype),
+        )

engines/dramabox/vendor/ltx2/ltx_core/components/patchifiers.py

@@ -0,0 +1,348 @@
+import math
+from typing import Optional, Tuple
+
+import einops
+import torch
+
+from ltx_core.components.protocols import Patchifier
+from ltx_core.types import AudioLatentShape, SpatioTemporalScaleFactors, VideoLatentShape
+
+
+class VideoLatentPatchifier(Patchifier):
+    def __init__(self, patch_size: int):
+        # Patch sizes for video latents.
+        self._patch_size = (
+            1,  # temporal dimension
+            patch_size,  # height dimension
+            patch_size,  # width dimension
+        )
+
+    @property
+    def patch_size(self) -> Tuple[int, int, int]:
+        return self._patch_size
+
+    def get_token_count(self, tgt_shape: VideoLatentShape) -> int:
+        return math.prod(tgt_shape.to_torch_shape()[2:]) // math.prod(self._patch_size)
+
+    def patchify(
+        self,
+        latents: torch.Tensor,
+    ) -> torch.Tensor:
+        latents = einops.rearrange(
+            latents,
+            "b c (f p1) (h p2) (w p3) -> b (f h w) (c p1 p2 p3)",
+            p1=self._patch_size[0],
+            p2=self._patch_size[1],
+            p3=self._patch_size[2],
+        )
+
+        return latents
+
+    def unpatchify(
+        self,
+        latents: torch.Tensor,
+        output_shape: VideoLatentShape,
+    ) -> torch.Tensor:
+        assert self._patch_size[0] == 1, "Temporal patch size must be 1 for symmetric patchifier"
+
+        patch_grid_frames = output_shape.frames // self._patch_size[0]
+        patch_grid_height = output_shape.height // self._patch_size[1]
+        patch_grid_width = output_shape.width // self._patch_size[2]
+
+        latents = einops.rearrange(
+            latents,
+            "b (f h w) (c p q) -> b c f (h p) (w q)",
+            f=patch_grid_frames,
+            h=patch_grid_height,
+            w=patch_grid_width,
+            p=self._patch_size[1],
+            q=self._patch_size[2],
+        )
+
+        return latents
+
+    def get_patch_grid_bounds(
+        self,
+        output_shape: AudioLatentShape | VideoLatentShape,
+        device: Optional[torch.device] = None,
+    ) -> torch.Tensor:
+        """
+        Return the per-dimension bounds [inclusive start, exclusive end) for every
+        patch produced by `patchify`. The bounds are expressed in the original
+        video grid coordinates: frame/time, height, and width.
+        The resulting tensor is shaped `[batch_size, 3, num_patches, 2]`, where:
+            - axis 1 (size 3) enumerates (frame/time, height, width) dimensions
+            - axis 3 (size 2) stores `[start, end)` indices within each dimension
+        Args:
+            output_shape: Video grid description containing frames, height, and width.
+            device: Device of the latent tensor.
+        """
+        if not isinstance(output_shape, VideoLatentShape):
+            raise ValueError("VideoLatentPatchifier expects VideoLatentShape when computing coordinates")
+
+        frames = output_shape.frames
+        height = output_shape.height
+        width = output_shape.width
+        batch_size = output_shape.batch
+
+        # Validate inputs to ensure positive dimensions
+        assert frames > 0, f"frames must be positive, got {frames}"
+        assert height > 0, f"height must be positive, got {height}"
+        assert width > 0, f"width must be positive, got {width}"
+        assert batch_size > 0, f"batch_size must be positive, got {batch_size}"
+
+        # Generate grid coordinates for each dimension (frame, height, width)
+        # We use torch.arange to create the starting coordinates for each patch.
+        # indexing='ij' ensures the dimensions are in the order (frame, height, width).
+        grid_coords = torch.meshgrid(
+            torch.arange(start=0, end=frames, step=self._patch_size[0], device=device),
+            torch.arange(start=0, end=height, step=self._patch_size[1], device=device),
+            torch.arange(start=0, end=width, step=self._patch_size[2], device=device),
+            indexing="ij",
+        )
+
+        # Stack the grid coordinates to create the start coordinates tensor.
+        # Shape becomes (3, grid_f, grid_h, grid_w)
+        patch_starts = torch.stack(grid_coords, dim=0)
+
+        # Create a tensor containing the size of a single patch:
+        # (frame_patch_size, height_patch_size, width_patch_size).
+        # Reshape to (3, 1, 1, 1) to enable broadcasting when adding to the start coordinates.
+        patch_size_delta = torch.tensor(
+            self._patch_size,
+            device=patch_starts.device,
+            dtype=patch_starts.dtype,
+        ).view(3, 1, 1, 1)
+
+        # Calculate end coordinates: start + patch_size
+        # Shape becomes (3, grid_f, grid_h, grid_w)
+        patch_ends = patch_starts + patch_size_delta
+
+        # Stack start and end coordinates together along the last dimension
+        # Shape becomes (3, grid_f, grid_h, grid_w, 2), where the last dimension is [start, end]
+        latent_coords = torch.stack((patch_starts, patch_ends), dim=-1)
+
+        # Broadcast to batch size and flatten all spatial/temporal dimensions into one sequence.
+        # Final Shape: (batch_size, 3, num_patches, 2)
+        latent_coords = einops.repeat(
+            latent_coords,
+            "c f h w bounds -> b c (f h w) bounds",
+            b=batch_size,
+            bounds=2,
+        )
+
+        return latent_coords
+
+
+def get_pixel_coords(
+    latent_coords: torch.Tensor,
+    scale_factors: SpatioTemporalScaleFactors,
+    causal_fix: bool = False,
+) -> torch.Tensor:
+    """
+    Map latent-space `[start, end)` coordinates to their pixel-space equivalents by scaling
+    each axis (frame/time, height, width) with the corresponding VAE downsampling factors.
+    Optionally compensate for causal encoding that keeps the first frame at unit temporal scale.
+    Args:
+        latent_coords: Tensor of latent bounds shaped `(batch, 3, num_patches, 2)`.
+        scale_factors: SpatioTemporalScaleFactors tuple `(temporal, height, width)` with integer scale factors applied
+        per axis.
+        causal_fix: When True, rewrites the temporal axis of the first frame so causal VAEs
+            that treat frame zero differently still yield non-negative timestamps.
+    """
+    # Broadcast the VAE scale factors so they align with the `(batch, axis, patch, bound)` layout.
+    broadcast_shape = [1] * latent_coords.ndim
+    broadcast_shape[1] = -1  # axis dimension corresponds to (frame/time, height, width)
+    scale_tensor = torch.tensor(scale_factors, device=latent_coords.device).view(*broadcast_shape)
+
+    # Apply per-axis scaling to convert latent bounds into pixel-space coordinates.
+    pixel_coords = latent_coords * scale_tensor
+
+    if causal_fix:
+        # VAE temporal stride for the very first frame is 1 instead of `scale_factors[0]`.
+        # Shift and clamp to keep the first-frame timestamps causal and non-negative.
+        pixel_coords[:, 0, ...] = (pixel_coords[:, 0, ...] + 1 - scale_factors[0]).clamp(min=0)
+
+    return pixel_coords
+
+
+class AudioPatchifier(Patchifier):
+    def __init__(
+        self,
+        patch_size: int,
+        sample_rate: int = 16000,
+        hop_length: int = 160,
+        audio_latent_downsample_factor: int = 4,
+        is_causal: bool = True,
+        shift: int = 0,
+    ):
+        """
+        Patchifier tailored for spectrogram/audio latents.
+        Args:
+            patch_size: Number of mel bins combined into a single patch. This
+                controls the resolution along the frequency axis.
+            sample_rate: Original waveform sampling rate. Used to map latent
+                indices back to seconds so downstream consumers can align audio
+                and video cues.
+            hop_length: Window hop length used for the spectrogram. Determines
+                how many real-time samples separate two consecutive latent frames.
+            audio_latent_downsample_factor: Ratio between spectrogram frames and
+                latent frames; compensates for additional downsampling inside the
+                VAE encoder.
+            is_causal: When True, timing is shifted to account for causal
+                receptive fields so timestamps do not peek into the future.
+            shift: Integer offset applied to the latent indices. Enables
+                constructing overlapping windows from the same latent sequence.
+        """
+        self.hop_length = hop_length
+        self.sample_rate = sample_rate
+        self.audio_latent_downsample_factor = audio_latent_downsample_factor
+        self.is_causal = is_causal
+        self.shift = shift
+        self._patch_size = (1, patch_size, patch_size)
+
+    @property
+    def patch_size(self) -> Tuple[int, int, int]:
+        return self._patch_size
+
+    def get_token_count(self, tgt_shape: AudioLatentShape) -> int:
+        return tgt_shape.frames
+
+    def _get_audio_latent_time_in_sec(
+        self,
+        start_latent: int,
+        end_latent: int,
+        dtype: torch.dtype,
+        device: Optional[torch.device] = None,
+    ) -> torch.Tensor:
+        """
+        Converts latent indices into real-time seconds while honoring causal
+        offsets and the configured hop length.
+        Args:
+            start_latent: Inclusive start index inside the latent sequence. This
+                sets the first timestamp returned.
+            end_latent: Exclusive end index. Determines how many timestamps get
+                generated.
+            dtype: Floating-point dtype used for the returned tensor, allowing
+                callers to control precision.
+            device: Target device for the timestamp tensor. When omitted the
+                computation occurs on CPU to avoid surprising GPU allocations.
+        """
+        if device is None:
+            device = torch.device("cpu")
+
+        audio_latent_frame = torch.arange(start_latent, end_latent, dtype=dtype, device=device)
+
+        audio_mel_frame = audio_latent_frame * self.audio_latent_downsample_factor
+
+        if self.is_causal:
+            # Frame offset for causal alignment.
+            # The "+1" ensures the timestamp corresponds to the first sample that is fully available.
+            causal_offset = 1
+            audio_mel_frame = (audio_mel_frame + causal_offset - self.audio_latent_downsample_factor).clip(min=0)
+
+        return audio_mel_frame * self.hop_length / self.sample_rate
+
+    def _compute_audio_timings(
+        self,
+        batch_size: int,
+        num_steps: int,
+        device: Optional[torch.device] = None,
+    ) -> torch.Tensor:
+        """
+        Builds a `(B, 1, T, 2)` tensor containing timestamps for each latent frame.
+        This helper method underpins `get_patch_grid_bounds` for the audio patchifier.
+        Args:
+            batch_size: Number of sequences to broadcast the timings over.
+            num_steps: Number of latent frames (time steps) to convert into timestamps.
+            device: Device on which the resulting tensor should reside.
+        """
+        resolved_device = device
+        if resolved_device is None:
+            resolved_device = torch.device("cpu")
+
+        start_timings = self._get_audio_latent_time_in_sec(
+            self.shift,
+            num_steps + self.shift,
+            torch.float32,
+            resolved_device,
+        )
+        start_timings = start_timings.unsqueeze(0).expand(batch_size, -1).unsqueeze(1)
+
+        end_timings = self._get_audio_latent_time_in_sec(
+            self.shift + 1,
+            num_steps + self.shift + 1,
+            torch.float32,
+            resolved_device,
+        )
+        end_timings = end_timings.unsqueeze(0).expand(batch_size, -1).unsqueeze(1)
+
+        return torch.stack([start_timings, end_timings], dim=-1)
+
+    def patchify(
+        self,
+        audio_latents: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Flattens the audio latent tensor along time. Use `get_patch_grid_bounds`
+        to derive timestamps for each latent frame based on the configured hop
+        length and downsampling.
+        Args:
+            audio_latents: Latent tensor to patchify.
+        Returns:
+            Flattened patch tokens tensor. Use `get_patch_grid_bounds` to compute the
+            corresponding timing metadata when needed.
+        """
+        audio_latents = einops.rearrange(
+            audio_latents,
+            "b c t f -> b t (c f)",
+        )
+
+        return audio_latents
+
+    def unpatchify(
+        self,
+        audio_latents: torch.Tensor,
+        output_shape: AudioLatentShape,
+    ) -> torch.Tensor:
+        """
+        Restores the `(B, C, T, F)` spectrogram tensor from flattened patches.
+        Use `get_patch_grid_bounds` to recompute the timestamps that describe each
+        frame's position in real time.
+        Args:
+            audio_latents: Latent tensor to unpatchify.
+            output_shape: Shape of the unpatched output tensor.
+        Returns:
+            Unpatched latent tensor. Use `get_patch_grid_bounds` to compute the timing
+            metadata associated with the restored latents.
+        """
+        # audio_latents shape: (batch, time, freq * channels)
+        audio_latents = einops.rearrange(
+            audio_latents,
+            "b t (c f) -> b c t f",
+            c=output_shape.channels,
+            f=output_shape.mel_bins,
+        )
+
+        return audio_latents
+
+    def get_patch_grid_bounds(
+        self,
+        output_shape: AudioLatentShape | VideoLatentShape,
+        device: Optional[torch.device] = None,
+    ) -> torch.Tensor:
+        """
+        Return the temporal bounds `[inclusive start, exclusive end)` for every
+        patch emitted by `patchify`. For audio this corresponds to timestamps in
+        seconds aligned with the original spectrogram grid.
+        The returned tensor has shape `[batch_size, 1, time_steps, 2]`, where:
+            - axis 1 (size 1) represents the temporal dimension
+            - axis 3 (size 2) stores the `[start, end)` timestamps per patch
+        Args:
+            output_shape: Audio grid specification describing the number of time steps.
+            device: Target device for the returned tensor.
+        """
+        if not isinstance(output_shape, AudioLatentShape):
+            raise ValueError("AudioPatchifier expects AudioLatentShape when computing coordinates")
+
+        return self._compute_audio_timings(output_shape.batch, output_shape.frames, device)

engines/dramabox/vendor/ltx2/ltx_core/components/protocols.py

@@ -0,0 +1,101 @@
+from typing import Protocol, Tuple
+
+import torch
+
+from ltx_core.types import AudioLatentShape, VideoLatentShape
+
+
+class Patchifier(Protocol):
+    """
+    Protocol for patchifiers that convert latent tensors into patches and assemble them back.
+    """
+
+    def patchify(
+        self,
+        latents: torch.Tensor,
+    ) -> torch.Tensor:
+        ...
+        """
+        Convert latent tensors into flattened patch tokens.
+        Args:
+            latents: Latent tensor to patchify.
+        Returns:
+            Flattened patch tokens tensor.
+        """
+
+    def unpatchify(
+        self,
+        latents: torch.Tensor,
+        output_shape: AudioLatentShape | VideoLatentShape,
+    ) -> torch.Tensor:
+        """
+        Converts latent tensors between spatio-temporal formats and flattened sequence representations.
+        Args:
+            latents: Patch tokens that must be rearranged back into the latent grid constructed by `patchify`.
+            output_shape: Shape of the output tensor. Note that output_shape is either AudioLatentShape or
+            VideoLatentShape.
+        Returns:
+            Dense latent tensor restored from the flattened representation.
+        """
+
+    @property
+    def patch_size(self) -> Tuple[int, int, int]:
+        ...
+        """
+        Returns the patch size as a tuple of (temporal, height, width) dimensions
+        """
+
+    def get_patch_grid_bounds(
+        self,
+        output_shape: AudioLatentShape | VideoLatentShape,
+        device: torch.device | None = None,
+    ) -> torch.Tensor:
+        ...
+        """
+        Compute metadata describing where each latent patch resides within the
+        grid specified by `output_shape`.
+        Args:
+            output_shape: Target grid layout for the patches.
+            device: Target device for the returned tensor.
+        Returns:
+            Tensor containing patch coordinate metadata such as spatial or temporal intervals.
+        """
+
+
+class SchedulerProtocol(Protocol):
+    """
+    Protocol for schedulers that provide a sigmas schedule tensor for a
+    given number of steps. Device is cpu.
+    """
+
+    def execute(self, steps: int, **kwargs) -> torch.FloatTensor: ...
+
+
+class GuiderProtocol(Protocol):
+    """
+    Protocol for guiders that compute a delta tensor given conditioning inputs.
+    The returned delta should be added to the conditional output (cond), enabling
+    multiple guiders to be chained together by accumulating their deltas.
+    """
+
+    scale: float
+
+    def delta(self, cond: torch.Tensor, uncond: torch.Tensor) -> torch.Tensor: ...
+
+    def enabled(self) -> bool:
+        """
+        Returns whether the corresponding perturbation is enabled. E.g. for CFG, this should return False if the scale
+        is 1.0.
+        """
+        ...
+
+
+class DiffusionStepProtocol(Protocol):
+    """
+    Protocol for diffusion steps that provide a next sample tensor for a given current sample tensor,
+    current denoised sample tensor, and sigmas tensor.
+    """
+
+    def step(
+        self, sample: torch.Tensor, denoised_sample: torch.Tensor, sigmas: torch.Tensor, step_index: int, **kwargs
+    ) -> torch.Tensor: ...

engines/dramabox/vendor/ltx2/ltx_core/components/schedulers.py

@@ -0,0 +1,130 @@
+import math
+from functools import lru_cache
+
+import numpy
+import scipy
+import torch
+
+from ltx_core.components.protocols import SchedulerProtocol
+
+BASE_SHIFT_ANCHOR = 1024
+MAX_SHIFT_ANCHOR = 4096
+
+
+class LTX2Scheduler(SchedulerProtocol):
+    """
+    Default scheduler for LTX-2 diffusion sampling.
+    Generates a sigma schedule with token-count-dependent shifting and optional
+    stretching to a terminal value.
+    """
+
+    def execute(
+        self,
+        steps: int,
+        latent: torch.Tensor | None = None,
+        max_shift: float = 2.05,
+        base_shift: float = 0.95,
+        stretch: bool = True,
+        terminal: float = 0.1,
+        default_number_of_tokens: int = MAX_SHIFT_ANCHOR,
+        **_kwargs,
+    ) -> torch.FloatTensor:
+        tokens = math.prod(latent.shape[2:]) if latent is not None else default_number_of_tokens
+        sigmas = torch.linspace(1.0, 0.0, steps + 1)
+
+        x1 = BASE_SHIFT_ANCHOR
+        x2 = MAX_SHIFT_ANCHOR
+        mm = (max_shift - base_shift) / (x2 - x1)
+        b = base_shift - mm * x1
+        sigma_shift = (tokens) * mm + b
+
+        power = 1
+        sigmas = torch.where(
+            sigmas != 0,
+            math.exp(sigma_shift) / (math.exp(sigma_shift) + (1 / sigmas - 1) ** power),
+            0,
+        )
+
+        # Stretch sigmas so that its final value matches the given terminal value.
+        if stretch:
+            non_zero_mask = sigmas != 0
+            non_zero_sigmas = sigmas[non_zero_mask]
+            one_minus_z = 1.0 - non_zero_sigmas
+            scale_factor = one_minus_z[-1] / (1.0 - terminal)
+            stretched = 1.0 - (one_minus_z / scale_factor)
+            sigmas[non_zero_mask] = stretched
+
+        return sigmas.to(torch.float32)
+
+
+class LinearQuadraticScheduler(SchedulerProtocol):
+    """
+    Scheduler with linear steps followed by quadratic steps.
+    Produces a sigma schedule that transitions linearly up to a threshold,
+    then follows a quadratic curve for the remaining steps.
+    """
+
+    def execute(
+        self, steps: int, threshold_noise: float = 0.025, linear_steps: int | None = None, **_kwargs
+    ) -> torch.FloatTensor:
+        if steps == 1:
+            return torch.FloatTensor([1.0, 0.0])
+
+        if linear_steps is None:
+            linear_steps = steps // 2
+        linear_sigma_schedule = [i * threshold_noise / linear_steps for i in range(linear_steps)]
+        threshold_noise_step_diff = linear_steps - threshold_noise * steps
+        quadratic_steps = steps - linear_steps
+        quadratic_sigma_schedule = []
+        if quadratic_steps > 0:
+            quadratic_coef = threshold_noise_step_diff / (linear_steps * quadratic_steps**2)
+            linear_coef = threshold_noise / linear_steps - 2 * threshold_noise_step_diff / (quadratic_steps**2)
+            const = quadratic_coef * (linear_steps**2)
+            quadratic_sigma_schedule = [
+                quadratic_coef * (i**2) + linear_coef * i + const for i in range(linear_steps, steps)
+            ]
+        sigma_schedule = linear_sigma_schedule + quadratic_sigma_schedule + [1.0]
+        sigma_schedule = [1.0 - x for x in sigma_schedule]
+        return torch.FloatTensor(sigma_schedule)
+
+
+class BetaScheduler(SchedulerProtocol):
+    """
+    Scheduler using a beta distribution to sample timesteps.
+    Based on: https://arxiv.org/abs/2407.12173
+    """
+
+    shift = 2.37
+    timesteps_length = 10000
+
+    def execute(self, steps: int, alpha: float = 0.6, beta: float = 0.6) -> torch.FloatTensor:
+        """
+        Execute the beta scheduler.
+        Args:
+            steps: The number of steps to execute the scheduler for.
+            alpha: The alpha parameter for the beta distribution.
+            beta: The beta parameter for the beta distribution.
+        Warnings:
+            The number of steps within `sigmas` theoretically might be less than `steps+1`,
+            because of the deduplication of the identical timesteps
+        Returns:
+            A tensor of sigmas.
+        """
+        model_sampling_sigmas = _precalculate_model_sampling_sigmas(self.shift, self.timesteps_length)
+        total_timesteps = len(model_sampling_sigmas) - 1
+        ts = 1 - numpy.linspace(0, 1, steps, endpoint=False)
+        ts = numpy.rint(scipy.stats.beta.ppf(ts, alpha, beta) * total_timesteps).tolist()
+        ts = list(dict.fromkeys(ts))
+
+        sigmas = [float(model_sampling_sigmas[int(t)]) for t in ts] + [0.0]
+        return torch.FloatTensor(sigmas)
+
+
+@lru_cache(maxsize=5)
+def _precalculate_model_sampling_sigmas(shift: float, timesteps_length: int) -> torch.Tensor:
+    timesteps = torch.arange(1, timesteps_length + 1, 1) / timesteps_length
+    return torch.Tensor([flux_time_shift(shift, 1.0, t) for t in timesteps])
+
+
+def flux_time_shift(mu: float, sigma: float, t: float) -> float:
+    return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)

engines/dramabox/vendor/ltx2/ltx_core/conditioning/__init__.py

@@ -0,0 +1,19 @@
+"""Conditioning utilities: latent state, tools, and conditioning types."""
+
+from ltx_core.conditioning.exceptions import ConditioningError
+from ltx_core.conditioning.item import ConditioningItem
+from ltx_core.conditioning.types import (
+    ConditioningItemAttentionStrengthWrapper,
+    VideoConditionByKeyframeIndex,
+    VideoConditionByLatentIndex,
+    VideoConditionByReferenceLatent,
+)
+
+__all__ = [
+    "ConditioningError",
+    "ConditioningItem",
+    "ConditioningItemAttentionStrengthWrapper",
+    "VideoConditionByKeyframeIndex",
+    "VideoConditionByLatentIndex",
+    "VideoConditionByReferenceLatent",
+]

engines/dramabox/vendor/ltx2/ltx_core/conditioning/exceptions.py

@@ -0,0 +1,4 @@
+class ConditioningError(Exception):
+    """
+    Class for conditioning-related errors.
+    """

engines/dramabox/vendor/ltx2/ltx_core/conditioning/item.py

@@ -0,0 +1,20 @@
+from typing import Protocol
+
+from ltx_core.tools import LatentTools
+from ltx_core.types import LatentState
+
+
+class ConditioningItem(Protocol):
+    """Protocol for conditioning items that modify latent state during diffusion."""
+
+    def apply_to(self, latent_state: LatentState, latent_tools: LatentTools) -> LatentState:
+        """
+        Apply the conditioning to the latent state.
+        Args:
+            latent_state: The latent state to apply the conditioning to. This is state always patchified.
+        Returns:
+            The latent state after the conditioning has been applied.
+        IMPORTANT: If the conditioning needs to add extra tokens to the latent, it should add them to the end of the
+        latent.
+        """
+        ...

engines/dramabox/vendor/ltx2/ltx_core/conditioning/mask_utils.py

@@ -0,0 +1,210 @@
+"""Utilities for building 2D self-attention masks for conditioning items."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import torch
+
+if TYPE_CHECKING:
+    from ltx_core.types import LatentState
+
+
+def resolve_cross_mask(
+    attention_mask: float | int | torch.Tensor,
+    num_new_tokens: int,
+    batch_size: int,
+    device: torch.device,
+    dtype: torch.dtype,
+) -> torch.Tensor:
+    """Convert an attention_mask (scalar or tensor) to a (B, M) cross_mask tensor.
+    Args:
+        attention_mask: Scalar value applied uniformly, 1D tensor of shape (M,)
+            broadcast across batch, or 2D tensor of shape (B, M).
+        num_new_tokens: Number of new conditioning tokens M.
+        batch_size: Batch size B.
+        device: Device for the output tensor.
+        dtype: Data type for the output tensor.
+    Returns:
+        Cross-mask tensor of shape (B, M).
+    """
+    if isinstance(attention_mask, (int, float)):
+        return torch.full(
+            (batch_size, num_new_tokens),
+            fill_value=float(attention_mask),
+            device=device,
+            dtype=dtype,
+        )
+    mask = attention_mask.to(device=device, dtype=dtype)
+
+    # Handle scalar (0-D) tensor like a Python scalar.
+    if mask.dim() == 0:
+        return torch.full(
+            (batch_size, num_new_tokens),
+            fill_value=float(mask.item()),
+            device=device,
+            dtype=dtype,
+        )
+
+    if mask.dim() == 1:
+        if mask.shape[0] != num_new_tokens:
+            raise ValueError(
+                f"1-D attention_mask length must equal num_new_tokens ({num_new_tokens}), got shape {tuple(mask.shape)}"
+            )
+        mask = mask.unsqueeze(0).expand(batch_size, -1)
+    elif mask.dim() == 2:
+        b, m = mask.shape
+        if m != num_new_tokens:
+            raise ValueError(
+                f"2-D attention_mask second dimension must equal num_new_tokens ({num_new_tokens}), "
+                f"got shape {tuple(mask.shape)}"
+            )
+        if b not in (batch_size, 1):
+            raise ValueError(
+                f"2-D attention_mask batch dimension must equal batch_size ({batch_size}) or 1, "
+                f"got shape {tuple(mask.shape)}"
+            )
+        if b == 1 and batch_size > 1:
+            mask = mask.expand(batch_size, -1)
+    else:
+        raise ValueError(
+            f"attention_mask tensor must be 0-D, 1-D, or 2-D, got {mask.dim()}-D with shape {tuple(mask.shape)}"
+        )
+    return mask
+
+
+def update_attention_mask(
+    latent_state: LatentState,
+    attention_mask: float | torch.Tensor | None,
+    num_noisy_tokens: int,
+    num_new_tokens: int,
+    batch_size: int,
+    device: torch.device,
+    dtype: torch.dtype,
+) -> torch.Tensor | None:
+    """Build or update the self-attention mask for newly appended conditioning tokens.
+    If *attention_mask* is ``None`` and no existing mask is present, returns
+    ``None``.  If *attention_mask* is ``None`` but an existing mask is present,
+    the mask is expanded with full attention (1s) for the new tokens so that
+    its dimensions stay consistent with the growing latent sequence.  Otherwise,
+    resolves *attention_mask* to a per-token cross-mask and expands the 2-D
+    attention mask via :func:`build_attention_mask`.
+    Args:
+        latent_state: Current latent state (provides the existing mask and total
+            existing-token count).
+        attention_mask: Per-token attention weight. Scalar, 1-D ``(M,)``, 2-D
+            ``(B, M)`` tensor, or ``None`` (no-op).
+        num_noisy_tokens: Number of original noisy tokens (from
+            ``latent_tools.target_shape.token_count()``).
+        num_new_tokens: Number of new conditioning tokens being appended.
+        batch_size: Batch size.
+        device: Device for the output tensor.
+        dtype: Data type for the output tensor.
+    Returns:
+        Updated attention mask of shape ``(B, N+M, N+M)``, or ``None`` if no
+        masking is needed.
+    """
+    if attention_mask is None:
+        if latent_state.attention_mask is None:
+            return None
+        # Existing mask present but no new mask requested: pad with 1s (full
+        # attention) so the mask dimensions stay consistent with the growing
+        # latent sequence.
+        cross_mask = torch.ones(batch_size, num_new_tokens, device=device, dtype=dtype)
+        return build_attention_mask(
+            existing_mask=latent_state.attention_mask,
+            num_noisy_tokens=num_noisy_tokens,
+            num_new_tokens=num_new_tokens,
+            num_existing_tokens=latent_state.latent.shape[1],
+            cross_mask=cross_mask,
+            device=device,
+            dtype=dtype,
+        )
+
+    cross_mask = resolve_cross_mask(attention_mask, num_new_tokens, batch_size, device, dtype)
+    return build_attention_mask(
+        existing_mask=latent_state.attention_mask,
+        num_noisy_tokens=num_noisy_tokens,
+        num_new_tokens=num_new_tokens,
+        num_existing_tokens=latent_state.latent.shape[1],
+        cross_mask=cross_mask,
+        device=device,
+        dtype=dtype,
+    )
+
+
+def build_attention_mask(
+    existing_mask: torch.Tensor | None,
+    num_noisy_tokens: int,
+    num_new_tokens: int,
+    num_existing_tokens: int,
+    cross_mask: torch.Tensor,
+    device: torch.device,
+    dtype: torch.dtype,
+) -> torch.Tensor:
+    """
+    Expand the attention mask to include newly appended conditioning tokens.
+    Each conditioning item appends M new reference tokens to the sequence. This function
+    builds a (B, N+M, N+M) attention mask with the following block structure:
+                     noisy      prev_ref    new_ref
+                   (N_noisy)   (N-N_noisy)    (M)
+                 ┌───────────┬───────────┬───────────┐
+        noisy    │           │           │           │
+       (N_noisy) │  existing │  existing │   cross   │
+                 │           │           │           │
+                 ├───────────┼───────────┼───────────┤
+       prev_ref  │           │           │           │
+      (N-N_noisy)│  existing │  existing │     0     │
+                 │           │           │           │
+                 ├───────────┼───────────┼───────────┤
+       new_ref   │           │           │           │
+         (M)     │   cross   │     0     │     1     │
+                 │           │           │           │
+                 └───────────┴───────────┴───────────┘
+    Where:
+      - **existing**: preserved from the previous mask (or 1.0 if first conditioning)
+      - **cross**: values from *cross_mask* (shape B, M), in [0, 1]
+      - **0**: no attention between different reference groups
+    Args:
+        existing_mask: Current attention mask of shape (B, N, N), or None if no mask exists yet.
+            When None, the top-left NxN block is filled with 1s (full attention between all
+            existing tokens including any prior reference tokens that had no mask).
+        num_noisy_tokens: Number of original noisy tokens (always at positions [0:num_noisy_tokens]).
+        num_new_tokens: Number of new conditioning tokens M being appended.
+        num_existing_tokens: Total number of current tokens N (noisy + any prior conditioning tokens).
+        cross_mask: Per-token attention weight of shape (B, M) controlling attention between
+            new reference tokens and noisy tokens. Values in [0, 1].
+        device: Device for the output tensor.
+        dtype: Data type for the output tensor.
+    Returns:
+        Attention mask of shape (B, N+M, N+M) with values in [0, 1].
+    """
+    batch_size = cross_mask.shape[0]
+    total = num_existing_tokens + num_new_tokens
+
+    # Start with zeros
+    mask = torch.zeros((batch_size, total, total), device=device, dtype=dtype)
+
+    # Top-left: preserve existing mask or fill with 1s for noisy tokens
+    if existing_mask is not None:
+        mask[:, :num_existing_tokens, :num_existing_tokens] = existing_mask
+    else:
+        mask[:, :num_existing_tokens, :num_existing_tokens] = 1.0
+
+    # Bottom-right: new reference tokens fully attend to themselves
+    mask[:, num_existing_tokens:, num_existing_tokens:] = 1.0
+
+    # Cross-attention between noisy tokens and new reference tokens
+    # cross_mask shape: (B, M) -> broadcast to (B, N_noisy, M) and (B, M, N_noisy)
+
+    # Noisy tokens attending to new reference tokens: [0:N_noisy, N:N+M]
+    # Each column j in this block gets cross_mask[:, j]
+    mask[:, :num_noisy_tokens, num_existing_tokens:] = cross_mask.unsqueeze(1)
+
+    # New reference tokens attending to noisy tokens: [N:N+M, 0:N_noisy]
+    # Each row i in this block gets cross_mask[:, i]
+    mask[:, num_existing_tokens:, :num_noisy_tokens] = cross_mask.unsqueeze(2)
+
+    # [N_noisy:N, N:N+M] and [N:N+M, N_noisy:N] remain 0 (no cross-ref attention)
+
+    return mask

engines/dramabox/vendor/ltx2/ltx_core/conditioning/types/__init__.py

@@ -0,0 +1,13 @@
+"""Conditioning type implementations."""
+
+from ltx_core.conditioning.types.attention_strength_wrapper import ConditioningItemAttentionStrengthWrapper
+from ltx_core.conditioning.types.keyframe_cond import VideoConditionByKeyframeIndex
+from ltx_core.conditioning.types.latent_cond import VideoConditionByLatentIndex
+from ltx_core.conditioning.types.reference_video_cond import VideoConditionByReferenceLatent
+
+__all__ = [
+    "ConditioningItemAttentionStrengthWrapper",
+    "VideoConditionByKeyframeIndex",
+    "VideoConditionByLatentIndex",
+    "VideoConditionByReferenceLatent",
+]

engines/dramabox/vendor/ltx2/ltx_core/conditioning/types/attention_strength_wrapper.py

@@ -0,0 +1,71 @@
+"""Wrapper conditioning item that adds attention masking to any inner conditioning."""
+
+from dataclasses import replace
+
+import torch
+
+from ltx_core.conditioning.item import ConditioningItem
+from ltx_core.conditioning.mask_utils import update_attention_mask
+from ltx_core.tools import LatentTools
+from ltx_core.types import LatentState
+
+
+class ConditioningItemAttentionStrengthWrapper(ConditioningItem):
+    """Wraps a conditioning item to add an attention mask for its tokens.
+    Separates the *attention-masking* concern from the underlying conditioning
+    logic (token layout, positional encoding, denoise strength).  The inner
+    conditioning item appends tokens to the latent sequence as usual, and this
+    wrapper then builds or updates the self-attention mask so that the newly
+    added tokens interact with the noisy tokens according to *attention_mask*.
+    Args:
+        conditioning: Any conditioning item that appends tokens to the latent.
+        attention_mask: Per-token attention weight controlling how strongly the
+            new conditioning tokens attend to/from noisy tokens.  Can be a
+            scalar (float) applied uniformly, or a tensor of shape ``(B, M)``
+            for spatial control, where ``M = F * H * W`` is the number of
+            patchified conditioning tokens.  Values in ``[0, 1]``.
+    Example::
+        cond = ConditioningItemAttentionStrengthWrapper(
+            VideoConditionByReferenceLatent(latent=ref, strength=1.0),
+            attention_mask=0.5,
+        )
+        state = cond.apply_to(latent_state, latent_tools)
+    """
+
+    def __init__(
+        self,
+        conditioning: ConditioningItem,
+        attention_mask: float | torch.Tensor,
+    ):
+        self.conditioning = conditioning
+        self.attention_mask = attention_mask
+
+    def apply_to(
+        self,
+        latent_state: LatentState,
+        latent_tools: LatentTools,
+    ) -> LatentState:
+        """Apply inner conditioning, then build the attention mask for its tokens."""
+        # Snapshot the original state for mask building
+        original_state = latent_state
+
+        # Inner conditioning appends tokens (positions, denoise mask, etc.)
+        new_state = self.conditioning.apply_to(latent_state, latent_tools)
+
+        num_new_tokens = new_state.latent.shape[1] - original_state.latent.shape[1]
+        if num_new_tokens == 0:
+            return new_state
+
+        # Build the attention mask using the *original* state as the reference
+        # so that the block structure is computed correctly.
+        new_attention_mask = update_attention_mask(
+            latent_state=original_state,
+            attention_mask=self.attention_mask,
+            num_noisy_tokens=latent_tools.target_shape.token_count(),
+            num_new_tokens=num_new_tokens,
+            batch_size=new_state.latent.shape[0],
+            device=new_state.latent.device,
+            dtype=new_state.latent.dtype,
+        )
+
+        return replace(new_state, attention_mask=new_attention_mask)