Files

T

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

2026-08-11 23:42:44 -03:00

194 KiB

Raw Permalink Blame History

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[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

Add initial MOSS-TTS LoRA training support
Add initial MOSS-TTS 8B Delay LoRA training through the unified Model Training node
Add MOSS dataset staging, manifest building, preparation, and an example training workflow
Add 4-bit MOSS LoRA training option for lower-VRAM training on smaller GPUs

Fixed

Fix MOSS direct TTS inference so narrator fallback is not forced when no reference is provided

[4.26.5] - 2026-05-23

Added

Add regression coverage for Chinese subtitle formatting and English spacing

Fixed

Fix Text to SRT Builder Chinese punctuation and spacing
Fix Chinese/Japanese/Korean subtitles being spaced incorrectly
Fix Chinese punctuation appearing as standalone text in generated SRT output
Fix Chinese quote spacing such as ：“吃饭了吗？”

[4.26.4] - 2026-05-23

Added

Add MOSS LoRA loading support for MOSS-TTS
Add local MOSS LoRA selection from the managed TTS moss_tts/loras folder
Install MOSS LoRA adapters from Hugging Face directly into the local TTS model folder instead of hidden cache storage
Show active MOSS LoRA adapters in generation info output
Show real ComfyUI popup errors for MOSS LoRA/base model mismatches
Add documentation for the MOSS LoRA folder layout and usage

Fixed

Fail fast when a MOSS LoRA was trained for the wrong base model
Improve MOSS compatibility with newer LoRA adapter packages and installer upgrades

[4.26.3] - 2026-05-19

Added

Keep local model labels visible in the dropdown when models are already downloaded

Fixed

Fix MOSS model selection instability after switching variants
Prevent value mismatch issues when switching between 1.7B and 8B variants
Improve compatibility with older workflows using local:* model values
Keep native dialogue mode execution mapped to the correct TTSD model

[4.26.2] - 2026-05-18

Added

Prevent Qwen3-TTS from crashing when flash-attn is not fully installed
Keep auto attention mode working without requiring manual setting changes

Fixed

Fix Qwen3-TTS generation failure on Windows with flash-attn issues
Automatically fall back to SDPA attention for better compatibility

[4.26.1] - 2026-05-18

Fixed

Fix RVC voice conversion matrix shape error in RMVPE fallback
Fix trained RVC models still returning original audio after the mel-filter fallback is used
Correct fallback mel filter shape for RMVPE pitch extraction
Improve RVC compatibility on Python 3.13 systems where librosa utility imports fail

[4.26.0] - 2026-05-18

Added

Add MOSS-TTS engine family with native multi-speaker dialogue
Add MOSS Local 1.7B, Delay 8B, and TTSD 8B model support in unified TTS workflows
Add native dialogue mapping support for [Character], [S1]/[1], and Speaker N formats
Native TTSD now shows explicit errors for unsupported controls (pause tags, inline edit tags, per-segment parameter changes, or more than 5 speakers) instead of silently switching models

Changed

Improve SRT native behavior with stable speaker mapping and clear incompatibility feedback

[4.25.21] - 2026-05-09

Added

Support mel filter options used by RMVPE pitch extraction on Python 3.13 setups

Fixed

Fix RVC voice conversion failing during RMVPE pitch extraction
Fix trained RVC models still returning original audio after the librosa utility fallback
Preserve normal librosa behavior while improving fallback compatibility

[4.25.20] - 2026-05-07

Added

Preserve normal librosa behavior for systems where librosa is already working

Fixed

Fix RVC voice conversion crash on some Python 3.13 setups
Fix trained RVC models returning the original audio when librosa utility imports fail
Add safer RVC voice conversion fallback behavior for broken numba/librosa environments

[4.25.19] - 2026-05-06

Added

Prevent startup failure when loading RMVPE pitch extraction utilities
Add safer fallback behavior for librosa utility imports used by RVC

Fixed

Fix RVC dataset prep crash on some Python 3.13 setups
Improve compatibility for environments hitting numba/librosa import-time typing errors

[4.25.18] - 2026-04-17

Added

Add safer loading fallback for OGG and other compressed voice reference files

Fixed

Fix Echo-TTS voice references from Character Voices and OGG files
Fix Echo-TTS using Character Voices and dropdown voices when only the file path is available
Improve compatibility when direct audio decoding is unavailable in some Windows environments

[4.25.17] - 2026-04-14

Changed

Improve IndexTTS-2 startup logs by removing stale Python 3.12 VRAM warnings and extra QwenEmotion debug output
Update IndexTTS-2 QwenEmotion loading to avoid deprecated dtype warnings

Fixed

Fix IndexTTS-2 silent failures and noisy startup warnings
Fix IndexTTS-2 accidentally enabling torch.compile from legacy workflow values
Fix IndexTTS-2 runs returning silent audio instead of a real error when generation fails
Improve IndexTTS-2 error message when torch.compile is enabled without Triton on CUDA

[4.25.16] - 2026-04-14

Added

Restore the official Higgs Audio 2 download to a known-good v2 snapshot

Changed

Improve Higgs Audio loading behavior by reducing repeated managed download log noise

Fixed

Fix Higgs Audio 2 download compatibility with current transformers support
Fix newer incompatible Higgs metadata already downloaded from Hugging Face

[4.25.15] - 2026-04-13

Changed

Improve Higgs Audio support for newer upstream model releases

Fixed

Fix Higgs Audio loading with recent upstream tokenizer updates
Fix Higgs Audio tokenizer loading failure after the latest Hugging Face tokenizer metadata change
Keep compatibility with older Higgs downloads and existing tokenizer layouts

[4.25.14] - 2026-04-12

Changed

Improve RVC training behavior when FAISS is present but unusable

Fixed

Fix RVC training FAISS package repair on Linux
Fix RVC index building failing when Python imports a broken stub faiss package with no real APIs
Reinstall faiss-cpu more aggressively when the environment contains broken FAISS leftovers

[4.25.13] - 2026-04-12

Added

Resume should no longer fall back to a fresh run just because checkpoint loading is blocked by weights_only defaults

Changed

Improve reliability of resumed RVC training runs

Fixed

Fix RVC training resume checkpoint loading
Fix RVC resume failing on newer PyTorch versions when loading saved training checkpoints

[4.25.12] - 2026-04-12

Added

Add clearer Higgs config detection logs for troubleshooting

Fixed

Fix Higgs Audio compatibility with recent upstream model updates
Fix Higgs Audio SRT language-tag processing crash
Fix Higgs Audio model loading failure with the latest Hugging Face release

[4.25.11] - 2026-04-12

Added

Resume now reuses the latest compatible checkpointed RVC job for the same output name, dataset, and sample rate

Changed

Improve resume messaging when no compatible saved checkpoints exist

Fixed

Fix RVC training resume after config changes
Fix RVC resume failing just because training settings changed between runs

[4.25.10] - 2026-04-12

Added

Show clearer install warnings when faiss is present but unusable

Changed

Improve RVC training FAISS setup on Linux
Improve RVC training setup when Linux CUDA faiss packages install incorrectly

Fixed

Fall back to faiss-cpu automatically if the GPU faiss package is broken

[4.25.9] - 2026-04-11

Changed

Improve RVC model training dataset prep reliability

Fixed

Fix RVC model training failing silently when RMVPE pitch extraction files are missing
Prevent RVC model training from continuing with a broken voiced dataset
Improve RVC model training index build diagnostics when FAISS imports still fail

[4.25.8] - 2026-04-11

Added

Clarify FAISS installation guidance for RVC model training

Changed

Improve RVC model training behavior when FAISS index dependencies are unavailable

Fixed

Improve RVC model training dependency compatibility
Fix RVC model training failing when only the system ffmpeg CLI is installed
Avoid RVC installer dependency conflicts when installing faiss-cpu

[4.25.7] - 2026-04-11

Added

Restore the example RVC model training workflow in fresh setups

Fixed

Fix RVC training startup failures
Fix missing bundled RVC training runtime files in clean installs
Improve RVC training multiprocessing compatibility on Linux and ROCm systems

[4.25.6] - 2026-04-10

Fixed

Fix ChatterBox model loading failure on some Python 3.13 setups
Fix ChatterBox crashing before generation when optional watermark dependency imports a fragile librosa path
Improve compatibility by loading watermark support only when it is actually used

[4.25.5] - 2026-04-08

Added

Add find and replace tools directly inside the editor
Add replace all, match options, and keyboard shortcuts for faster text editing
Add collapsible left panel control while keeping sidebar resizing available

Changed

Improve Multiline TTS Tag Editor editing workflow
Improve left panel toggle behavior and visual polish during resize

[4.25.4] - 2026-04-08

Added

Keep inline edit tag and voice restoration settings working as before
Reduce noisy console spam from deprecated settings API usage

Fixed

Fix TTS Audio Suite settings deprecation warnings
Fix ComfyUI settings warnings shown in the browser console

[4.25.3] - 2026-04-08

Added

Redesign SRT Advanced Options interface
Clarify subtitle mode and language-profile behavior directly in the UI

Changed

Add a cleaner compact layout for subtitle policy controls
Improve preset editing with visible modified states and smarter locking

[4.25.2] - 2026-04-06

Fixed

Fix ONNX Runtime installer preserving GPU acceleration
Fix TTS Audio Suite reinstalling CPU ONNX Runtime over existing GPU setups
Improve Python 3.13 OpenSeeFace dependency handling without replacing working ONNX installs
Better compatibility with shared ComfyUI environments when other nodes use onnxruntime-gpu

[4.25.1] - 2026-04-05

Added

Move internal RVC training shim files into the engine package where they belong

Changed

Improve Windows RVC training stability without changing the user workflow

Fixed

Fix RVC training resume/runtime packaging on Windows
Fix RVC training child processes failing after the runtime shim cleanup

[4.25.0] - 2026-04-05

Added

Add integrated RVC model training workflow
Add dataset preparation, training config, and unified training nodes for RVC voice models
Add live RVC training dashboard with progress, ETA, and recent loss visualization
Add resume and continue-from-model support for longer RVC training sessions

Changed

Improve RVC model loading, index auto-detection, and training documentation

[4.24.20] - 2026-04-03

Changed

Improve F5-TTS fallback loading when automatic model download does not complete cleanly

Fixed

Address F5-TTS model loading issue
Fix F5-TTS failing to load bundled vocab files on some Windows custom-node installs
Prevent broken fallback paths that could stop F5-TTS Speech Editor workflows from starting

[4.24.19] - 2026-04-03

Fixed

Address Step Audio EditX Mac compatibility issue
Fix Step Audio EditX audio loading on fragile Mac and Python environments
Preserve existing Step Audio EditX audio preprocessing behavior while applying the compatibility fallback
Improve compatibility when EditX crashes with get_call_template errors before inference starts

[4.24.18] - 2026-04-03

Added

Add kugel-2 support in the VibeVoice engine

Changed

Improve VibeVoice and KugelAudio model support
Improve VibeVoice pause-tag logging so custom character switching shows split segments clearly
Update VibeVoice documentation and model source listings

Fixed

Fix VibeVoice Hindi model downloads by fetching tokenizer files from the correct source

[4.24.17] - 2026-04-01

Added

Add Library guide tabs for character tags, parameter tags, inline tags, and SRT editing

Changed

Improve Multiline TTS Tag Editor gutter spacing and Library browsing
Improve gutter spacing so line numbers use less wasted horizontal space
Improve guide readability by showing one reference section at a time inside the editor

[4.24.16] - 2026-04-01

Added

Keep gutter numbers aligned while scrolling and resizing long SRT files

Changed

Improve editor stability by using the rendered text rows instead of separate gutter measurements

Fixed

Fix Multiline TTS Tag Editor gutter alignment for wrapped subtitles
Fix line numbers drifting out of sync on wrapped subtitle text

[4.24.15] - 2026-04-01

Changed

Improve reliability when repeated pause tags appear across SRT cues

Fixed

Fix Multiline TTS Tag Editor tag highlighting with repeated subtitle tags
Fix some subtitle tags not highlighting correctly after editor updates
Fix internal marker artifacts leaking into the editor view

[4.24.14] - 2026-03-31

Added

Add merge controls for combining neighboring subtitle cues in the editor
Add split-at-caret support for breaking one subtitle into two timed cues
Add a dedicated Multiline TTS Tag Editor guide with SRT workflow notes

Changed

Improve Multiline TTS Tag Editor SRT editing and documentation
Improve in-editor and README documentation so the editor features are easier to discover

[4.24.13] - 2026-03-31

Added

Add draggable subtitle timing controls directly in the SRT editor
Keep adjacent subtitle gaps stable while adjusting linked timings
Keep the editor scrollbar from shifting text when focus changes

Changed

Improve Multiline TTS Tag Editor SRT timing editing and line alignment

Fixed

Fix line numbers drifting out of sync when long lines wrap

[4.24.12] - 2026-03-30

Added

Add SageAttention support for KugelAudio attention handling

Changed

Improve KugelAudio logging consistency with normal VibeVoice

Fixed

Fix VibeVoice KugelAudio engine settings being ignored
Fix KugelAudio 4-bit quantization not applying
Fix KugelAudio inference steps, top-p, and attention mode controls not working

[4.24.11] - 2026-03-29

Fixed

Fix Character Voices reload issues and CosyVoice model switching
Fix Character Voices failing to restore saved voices correctly after reopening workflows
Fix Character Voices preview player staying unavailable until reselecting the voice
Fix TTS generation still using the previous CosyVoice model after switching between 0.5B and 0.5B RL variants

[4.24.10] - 2026-03-28

Added

Add support for nested local layouts such as VibeVoice7b-low-vram/4bit

Changed

Improve VibeVoice quantized model loading stability

Fixed

Fix VibeVoice local custom model support
Fix local community and quantized VibeVoice folders not appearing in the model list
Fix VibeVoice startup import failures that could block generation until restart

[4.24.9] - 2026-03-28

Added

Expand SRT Advanced Options language profile coverage
Add heuristic defaults for Spanish, French, Italian, German, Dutch, Russian, Romanian, Indonesian, Malay, Turkish, Polish, Czech, Swedish, Danish, Finnish, and Greek

Changed

Improve Auto language matching so more ASR languages pick a sensible profile automatically
Update the SRT Advanced Options dropdown to prefill the new language-specific defaults

[4.24.8] - 2026-03-28

Added

Add reusable ASR timing JSON output so timings can be saved and reused without rerunning Qwen3 ASR

Changed

Improve Text to SRT Builder TTS-ready subtitle generation
Add TTS-ready subtitle mode for longer single-line cues that stop at better semantic boundaries
Add Portuguese (Brazil) heuristic defaults for cleaner subtitle chunking

Fixed

Fix awkward subtitle splits on connectors like de, da, do, na, and para

Removed

Remove the duplicate SRT Advanced Options node variant

[4.24.7] - 2026-03-27

Fixed

Fix CosyVoice3 narrator voice selection and interruption
Fix CosyVoice3 SRT and text generation using the wrong narrator reference in some workflows
Fix CosyVoice3 stop button handling during subtitle generation
Fix Qwen3-ASR stop button handling during long transcriptions
Improve connected Character Voices compatibility with unified narrator inputs

[4.24.6] - 2026-03-27

Added

Refresh Multiline TTS Tag Editor experience
Add redesigned editor layout with clearer controls and guided tabs
Expand inline restore guidance and tooltip help for easier use

Changed

Improve preset, history, and library workflows inside the node
Improve slider styling, tag highlighting, and overall editor polish

[4.24.5] - 2026-03-27

Changed

Better wrapper memory accounting for unload and reload paths

Fixed

Fix ComfyUI model wrapper crashes on newer ComfyUI versions
Fix RVC and CosyVoice3 workflows crashing during model unload
Improve shared TTS model compatibility with ComfyUI 0.18 memory management

[4.24.4] - 2026-03-19

Added

Reduce misleading startup warnings during background checks

Fixed

Fix background dependency checker false warnings
Fix TTS Audio Suite reporting some installed packages as missing
Improve dependency detection for heavy libraries like diffusers

[4.24.3] - 2026-03-18

Changed

Improve Portuguese language tag parsing for ChatterBox subtitle generation

Fixed

Fix ChatterBox SRT multiline subtitle handling and 23-Lang language tag switching
Fix Classic ChatterBox and ChatterBox Official 23-Lang skipping later lines in multiline SRT subtitles
Fix ChatterBox Official 23-Lang language-only tags like [pt:] not switching the tagged subtitle line

[4.24.2] - 2026-03-18

Added

Document control-tag-safe SRT building for TTS workflows

Changed

Improve Text to SRT Builder subtitle generation and documentation
Improve subtitle generation from edited transcripts and plain text workflows in the README

Fixed

Fix hyphenated words being formatted incorrectly in generated subtitles

[4.24.1] - 2026-03-17

Changed

Improve Multiline TTS Tag Editor undo, caret, and inline tag stability
Improve Multiline TTS Tag Editor keyboard undo and redo behavior so shortcuts now match the toolbar buttons

Fixed

Fix Multiline TTS Tag Editor caret jumps after undo, redo, toolbar inserts, and click repositioning in multiline subtitle content
Fix Multiline TTS Tag Editor inline tag deletion and broken tag editing from showing leaked EDIT_START or EDIT_END text in the editor
Fix Multiline TTS Tag Editor Ctrl or Cmd undo shortcuts so they edit text locally instead of undoing moved ComfyUI nodes

[4.24.0] - 2026-03-17

Added

Add modular ASR subtitle building and transcript cleanup tools
Add the standalone ASR Punctuation / Truecase helper node for low-punctuation transcripts such as Granite outputs
Add the new Text to SRT Builder workflow with separate subtitle construction controls and a modular ASR-to-SRT flow
Add updated Granite and Qwen3-ASR workflow examples, including the new Unified ✏️ ASR Transcribe + SRT Builder workflow

Changed

Improve subtitle readability with smarter cue splitting, better display balancing, and optional cue-end punctuation normalization
Improve README documentation and auto-generated model download tables to better reflect the current workflow setup

[4.23.0] - 2026-03-16

Added

Add Granite ASR as a second transcription engine in the unified ASR node
Add optional custom timestamps and SRT output for Granite via the reused Qwen aligner
Expose ASR translation target and prompt controls on Granite and Qwen engine nodes

Changed

Add Granite ASR engine and improve unified ASR workflows
Update README and engine comparison docs for the new ASR support

Fixed

Fix subtitle overflow so long lines split into new cues instead of losing words

[4.22.7] - 2026-03-13

Added

Keep audio loading and resampling working without extra startup overhead

Fixed

Fix Python 3.13 audio compatibility in IndexTTS-2 and ChatterBox Official 23-Lang
Fix Python 3.13 setups that could crash during reference voice processing
Improve compatibility with newer numba and librosa environments

[4.22.6] - 2026-03-13

Added

Restore the Voice Capture record button and proper input device dropdown
IndexTTS-2 works in our current Python 3.13 environment, so no engine changes were needed for now

Changed

Improve system default microphone selection and logging

Fixed

Fix Voice Capture sound initialization and input device handling
Fix Voice Capture blocking ComfyUI startup on some Windows audio setups

[4.22.5] - 2026-03-10

Added

Adjust ChatterBox Official 23-Lang Russian stress install flow
install the Russian stress support package during normal setup instead of mutating Python packages during generation
keep the large Russian dictionary as on-demand model data in the TTS folder
retain the safer Official 23-Lang Russian download path and clearer safetensors fallback messages

[4.22.4] - 2026-03-10

Added

Add lazy Russian stress setup so Official 23-Lang Russian no longer downloads large assets for every user during install
Download the Russian stress dictionary on demand into the normal TTS model folder

Fixed

Fix ChatterBox Official 23-Lang Russian stress support
Improve Russian generation setup on Python 3.13 by installing the patched stress-labeling dependency only when needed
Clarify Official 23-Lang safetensors fallback messages so working model folders no longer look broken

[4.22.3] - 2026-03-10

Fixed

Fix ChatterBox Official 23-Lang Python 3.13 generation failure
Fix Official 23-Lang failing after model download on Python 3.13 environments
Improve ChatterBox Official 23-Lang reference audio preprocessing compatibility

[4.22.2] - 2026-03-10

Fixed

Improve ChatterBox community model compatibility and Russian model handling
Fix ChatterBox local community models losing language-specific loading behavior
Fix stale ChatterBox model folders not redownloading newly added required files
Improve ChatterBox compatibility with community tokenizer variants
Remove the broken ChatterBox Russian classic Cyrillic experimental option from the dropdown

[4.22.1] - 2026-03-01

Added

Setting the env var has no effect once numba is already loaded by another node
Now forces numba to disable JIT via its internal config in all cases

Fixed

Fix Qwen3-TTS still crashing when NUMBA_DISABLE_JIT was already set
Fix voice cloning crash persisting even when NUMBA_DISABLE_JIT=1 was set in environment

[4.22.0] - 2026-03-01

Added

Add Echo-TTS voice cloning engine
New Echo-TTS engine: DiT-based voice cloning with reference audio support
Best quality at 30 seconds or less per generation; longer text auto-chunked
Force Speaker KV option controls speaker identity consistency across chunks
Full support for character switching, pause tags, and SRT subtitle timing
Auto-downloads models on first use (~7.1GB total into ComfyUI/models/TTS/)
Note: Echo-TTS weights are non-commercial (CC-BY-NC-SA license)

Fixed

Fix chunking not working when pause tags are present in Echo-TTS and Qwen3-TTS
Fix batch inline edit tags not applying correctly in SRT generation mode

[4.21.20] - 2026-02-28

Fixed

Fix Qwen3-TTS crash with numba 0.64.0 on certain hardware
Fix crash when using voice cloning on systems with numba 0.64.0 and NumPy 2.x
Error was silently ignored at startup so the incompatibility was never resolved
Startup now correctly detects and disables broken numba JIT compilation automatically

[4.21.19] - 2026-02-28

Added

All engines (Qwen3-TTS, IndexTTS-2, RVC, etc.) now correctly protected at startup

Fixed

Fix numba JIT compatibility test to correctly detect crashes on all affected hardware
Fix startup test that was not catching the Numba/librosa crash on some systems

[4.21.18] - 2026-02-28

Added

Engines like IndexTTS-2, Qwen3-TTS, RVC and others failed with AttributeError

Changed

No performance impact for users on NumPy 1.x

Fixed

Fix all engines crashing with Numba/librosa JIT error on NumPy 2.x
Fix crash affecting all TTS engines on NumPy 2.x with certain hardware
in librosa audio processing (get_call_template error)
Numba JIT compatibility is now correctly detected and disabled at startup

[4.21.17] - 2026-02-22

Added

Release v4.21.17

Fixed

Fix: Removed experimental Numba @guvectorize bypass that caused AttributeError crashes on fragile environments.

[4.21.16] - 2026-02-21

Added

Release v4.21.16

Fixed

Fix: Comprehensive fix for Numba @guvectorize crashes with Librosa on Python 3.12+ and NumPy 2.4+

[4.21.15] - 2026-02-21

Added

Release v4.21.15

Fixed

Fix: Librosa 0.11.0 compatibility fixes for missing utilities
Fix: ValueError unpacking error in IndexTTS-2

[4.21.14] - 2026-02-21

Added

Add in-node audio player with seek bar, scrub, and volume controls for selected Character Voices entries
Add quick voice preview playback for dropdown voices without running full TTS
Keep existing Character Voices workflows compatible with no input/output changes

Changed

Improve Character Voices voice preview controls

Fixed

Fix Character Voices source metadata reporting when direct input is used

[4.21.13] - 2026-02-21

Added

Add Egyptian Arabic (oddadmix) model support to ChatterBox 23L
Added new Egyptian Arabic community model 'oddadmix/chatterbox-egyptian-v0'

Fixed

Fixed 'Vocabulary Mismatch' errors by automatically handling official v2 tokenizer dependencies
Updated documentation tables and README with correct Egyptian flags and dialect notes

[4.21.12] - 2026-02-21

Fixed

Fix KaraFan FFmpeg detection crashes
Fix KaraFan node crashing when FFmpeg is not found in standard paths

[4.21.11] - 2026-02-20

Added

Resolved TypeError NoneType is not callable during ComfyUI GC and NameError for CHATTERBOX_VC_AVAILABLE in manager.py
Replaced lazy module evaluations using getattr with explicit getters in manager.py
Fixed LoadedModel wrapper initializations in IndexTTS and RVC to include .real_model assignment
Added fallback explicit weak-ref assignment in model_manager.py to prevent NoneType exceptions

[4.21.10] - 2026-02-20

Added

Fixed crashes occurring on macOS and CPU-only devices during inference

Fixed

Fix Mac/MPS compatibility issues
Improved device compatibility across Step Audio EditX, RVC Engine, and VibeVoice engines

[4.21.9] - 2026-02-20

Added

Drastically reduce extension startup time

Changed

Optimize internal model loading for better initialization performance

Fixed

Fix ComfyUI slowing down during startup by making heavy AI modules load only when first requested
Improve dependency checking to instantly verify missing packages without freezing ComfyUI

[4.21.8] - 2026-02-20

Added

Clarify default auto-download characters and optional index pack

Changed

Improve RVC character pack links to direct dataset folders

Fixed

Fix ChatterBox 23L and RVC model download references
Fix ChatterBox 23L Vietnamese (Viterbox) source link

[4.21.7] - 2026-02-20

Added

The RVC Vocal Removal node (using KaraFan) now correctly respects the system PATH or ComfyUI environmental variables when looking for ffmpeg.

Fixed

Fix RVC Vocal Remover failing due to hardcoded ffmpeg path
This resolves an issue where vocal removal would fail on specific portable Windows installations even if ffmpeg was installed.

[4.21.6] - 2026-02-19

Added

Add Qwen3 integration + ASR example workflow
Document ICL mode requires reference audio + transcript for best quality
Document X-Vector mode uses audio only, no transcript needed

Changed

Add Qwen3-TTS + ASR example workflow and improve Qwen3-TTS documentation

Fixed

Clarify that Qwen3-TTS auto-selects and downloads the correct model
Fix: style instructions are ignored in Base (voice cloning) model — now clearly noted

[4.21.5] - 2026-02-18

Added

Qwen3-TTS now correctly regenerates audio when model size changes

Fixed

Fix Qwen3-TTS audio cache not invalidating when switching model sizes
Fix switching between 0.6B and 1.7B models returning stale cached audio

[4.21.4] - 2026-02-18

Changed

Update Qwen3-TTS default settings
Enable Qwen3 forced aligner by default for better ASR accuracy out-of-the-box
Update tooltip to explain VRAM usage and model download implications

[4.21.3] - 2026-02-17

Added

Backward compatible with older ComfyUI versions

Fixed

Fix crashes when using RVC or IndexTTS alongside diffusion nodes on ComfyUI 0.13+
Fix AttributeError crash when ComfyUI memory manager runs after RVC voice conversion
Fix AttributeError crash when ComfyUI memory manager runs after IndexTTS generation
Fix affected workflows: any mix of TTS and image/video generation nodes (Flux, Wan, etc.)

[4.21.2] - 2026-02-14

Added

These engines auto-detect language from text and don't use language parameters

Changed

Console output is now cleaner when using VibeVoice or KugelAudio models

Fixed

Fix VibeVoice/KugelAudio console log spam

Removed

Remove misleading language auto-switching logs for VibeVoice and KugelAudio engines

[4.21.1] - 2026-02-13

Added

Ensure opt_narrator input is validated before overriding dropdown selection
Apply consistent fallback logic across all engines using Unified nodes

Fixed

Fix narrator fallback behavior in Unified TTS nodes
Fix issue where bypassing Character Voices node caused fallback to default 'David' voice

[4.21.0] - 2026-02-11

Added

Add KugelAudio-0-Open multilingual support via VibeVoice
New Model Support: Added full support for KugelAudio-0-Open localized model (~7GB)
Voice Cloning: Enable high-quality voice cloning with KugelAudio
Seamless Integration: Automatically detected via VibeVoice Engine node
Improved Robustness: Fixed import errors affecting some ComfyUI installations
Documentation: Updated all engine comparison tables and guides

Fixed

Bug Fixes: Resolved issue where model would speak 'Speaker 1:' tags

[4.20.21] - 2026-02-09

Added

Fixes remaining crashes on Mac (MPS) and CPU-only systems.

Fixed

Fix persistent 'Torch not compiled with CUDA' error in Step Audio EditX

Removed

Patched internal auto_model.py to remove hardcoded CUDA dependencies.

[4.20.20] - 2026-02-09

Changed

Improve device detection for non-CUDA systems

Fixed

Fix Mac compatibility for Step Audio EditX and ChatterBox VC
Fix 'Torch not compiled with CUDA' error on Mac (Step Audio EditX)
Fix serialization error when loading models on Mac (ChatterBox VC)

[4.20.19] - 2026-02-06

Added

Higgs Audio engine was incorrectly being passed to Step Audio EditX editor
Inline tags now properly load Step Audio EditX engine for editing operations

Fixed

Fix Higgs Audio SRT inline tag processing crash
Fix AttributeError when using inline edit tags with Higgs Audio in SRT mode

[4.20.18] - 2026-02-06

Added

Safe VRAM unloading now works properly - audio quality preserved
Simplified cleanup to only reset necessary flags, let engine handle recreation
Improved console output clarity - removed debug noise
Updated node tooltip to reflect safe CUDA graph management
CUDA graphs now auto-recreate properly without audio degradation

Fixed

Fix Higgs Audio CUDA graph crashes when clearing VRAM and reloading
Fix generation quality issues after model reload with CUDA graphs enabled

[4.20.17] - 2026-02-06

Added

Properly cleanup CUDA graph memory pools on unload
Allow safe model reloading while keeping CUDA graph optimizations

Fixed

Fix Qwen3-TTS Windows CUDA graph crashes on model reload
Fix crash when clearing VRAM with CUDA graphs enabled
Fix crash when switching between TTS Text and SRT nodes
Fix crash when changing model parameters after VRAM clear

[4.20.16] - 2026-02-05

Added

Speed up install when dependencies are already present
Reduce noisy reinstall checks in installer output

Fixed

Optimize installer dependency checks

[4.20.15] - 2026-02-05

Changed

Improve Windows embedded Python support

Fixed

Fix Step Audio EditX failing on Windows embedded Python
Fix distutils compatibility issue
TENTATIVE FIX - User testing required

[4.20.14] - 2026-02-05

Fixed

Fix ComfyUI compatibility for TTS engines
Fix crashes on ComfyUI 0.12+ when freeing memory with TTS models

[4.20.13] - 2026-02-04

Fixed

Fix onnxruntime-gpu installation failing in Docker environments
Fix GPU acceleration being disabled due to unnecessary CPU fallback installation
Improve compatibility with containerized environments (Docker, WSL)
Resolve issue where both onnxruntime and onnxruntime-gpu were installed, causing CPU-only operation

[4.20.12] - 2026-02-04

Added

Add advanced ASR SRT tuning options for edge cases

Changed

Improve ASR subtitles readability in ASR Transcribe
Improve subtitle grouping to avoid awkward splits in songs and questions
Improve ASR subtitle tooltips with clearer guidance

[4.20.11] - 2026-02-04

Fixed

Fix Qwen3-ASR compatibility with newer Transformers
Fix speech transcription crashes on transformers 4.57+

[4.20.10] - 2026-02-04

Added

Enable single-GPU inference on Windows AMD Radeon RX 9070 XT and similar GPUs

Fixed

Fix compatibility with Windows AMD ROCm 7.2 PyTorch builds
Fix all TTS engines crashing on Windows AMD systems with PyTorch 2.9.1+rocmsdk20260116
Resolve 'module torch.distributed has no attribute ReduceOp' errors
Automatically detect and patch incomplete torch.distributed implementations

[4.20.9] - 2026-02-04

Added

Step Audio EditX now works on Apple Silicon (MPS) and CPU

Changed

CUDA users unaffected (same performance and behavior)

Fixed

Add Step Audio EditX Mac (MPS) compatibility patch
Fix crash on Mac: 'Torch not compiled with CUDA enabled'
Centralized patch system in utils/compatibility/ for maintainability
Note: Tentative fix addressing issue #245 - requires Mac user testing

[4.20.8] - 2026-02-04

Added

Extend Qwen3-TTS torch.compile optimizations to Voice Designer
Voice Designer now inherits torch.compile settings from Qwen3-TTS Engine node
Enable via use_torch_compile parameter in connected Engine node
Supports all compile modes (default, reduce-overhead, max-autotune)

Changed

Significant performance improvement (~2x speedup: 4 it/s → 8.3 it/s confirmed)

[4.20.7] - 2026-02-04

Added

Credit: Timing improvement contributed by @rongsam

Changed

Improve SRT output to show actual audio duration

Fixed

Fix SRT timing accuracy for video editors
Fix timing drift in stretch_to_fit and pad_with_silence modes
Better compatibility with video editing software

[4.20.6] - 2026-02-03

Added

Add Refresh Voice Cache node for dynamic voice workflows
New ♻️ Refresh Voice Cache node for on-the-fly voice discovery
Support dual signal passthrough to guarantee execution order
Enable dynamic workflows where voices are created during execution

Fixed

Fix voice files created mid-workflow being unavailable to downstream nodes

[4.20.5] - 2026-02-03

Added

Changing torch.compile or cuda_graphs now properly reloads the model
Backward compatible - no changes needed in workflows

Fixed

Fix Qwen3-TTS torch.compile optimization not applying when toggled on
Fix optimization settings not taking effect after changing from off to on

[4.20.4] - 2026-02-03

Added

Backward compatible with ComfyUI 0.11.x and earlier

Fixed

Fix ComfyUI 0.12.0+ compatibility issue
Fix AttributeError: 'Qwen3TTSModel' object has no attribute 'is_dynamic'
Compatibility fix affects all TTS engines using ComfyUI model management

[4.20.3] - 2026-02-03

Changed

Improve FunASR model loading reliability

Fixed

Fix Step Audio EditX failing to load on some systems
Fix 'NoneType' object is not callable error
Better compatibility with Python 3.12+
TENTATIVE FIX - User testing required

[4.20.2] - 2026-02-02

Added

Add guidance for Transformers 5.x incompatibility (Qwen3-TTS)
Add a startup warning when Transformers 5.x is installed
Cap supported Transformers versions to <=4.57.3 for Qwen3-TTS stability
Add a public report on the Qwen3-TTS tokenizer incompatibility

[4.20.1] - 2026-01-31

Added

Add .agent/ directory to .gitignore for personal development tools

Changed

Improve cache validation to check both voice description and reference text before loading from disk

Fixed

Fix Qwen3-TTS Voice Designer caching issues and add voice refinement support
Fix Voice Designer loading stale audio when reference text changes but description stays same
Add overwrite_character toggle to enable voice refinement workflow (default: False for backward compatibility)

[4.20.0] - 2026-01-31

Added

Add Qwen3-ASR transcription with unified SRT output
New ASR Transcribe node with text, SRT, and word timestamps
Built-in subtitle readability presets and punctuation-aware SRT formatting
Automatic Qwen3-ASR downloads with forced aligner support
Faster experimentation via ASR result caching

[4.19.12] - 2026-01-30

Added

Add Qwen3-TTS torch.compile optimizations for 1.7x speedup
Add torch.compile support with ~1.7x faster generation (5 it/s → 8.5 it/s on RTX 4090)
Add optimization mode selection: default (Windows), reduce-overhead (Linux), max-autotune (Linux)
Add optional manual CUDA graph capture (minimal gain over torch.compile alone)
Add comprehensive optimization guide with tested benchmark results and requirements
Note: Optimizations disabled by default, Windows users should use compile_mode='default'
Based on streaming implementation by dffdeeq: https://github.com/dffdeeq/Qwen3-TTS-streaming

Fixed

Improve compatibility with PyTorch 2.10+ and triton-windows 3.6+ required for optimizations
Fix encoding errors in startup messages

[4.19.11] - 2026-01-30

Added

Properly respect user's attention mechanism choice (sdpa/eager/flash_attention_2)

Fixed

Fix Qwen3-TTS Voice Designer attention selection on AMD GPUs
Fix flash_attn import error when user selects sdpa attention
Fix config parameter extraction in Voice Designer node

[4.19.10] - 2026-01-29

Added

Increase Qwen3-TTS max_tokens range to 64-8192 (previously limited to 512-4096)
Add clear warning when x_vector_only mode is enabled with reference text available

Fixed

Fix Qwen3-TTS mode selection and Character Voices reference text handling
Fix Character Voices always using folder reference text when character is selected from dropdown (manual field now only for direct audio)
Fix Qwen3-TTS x_vector_only mode properly respecting engine widget setting when reference text is available
Fix Qwen3-TTS progress bar showing accurate progress (0-100%) instead of tiny percentages

[4.19.9] - 2026-01-29

Added

Audio inputs without transcript now properly use x_vector_only mode (lower quality but functional)

Changed

Update warning messages to clarify engine-specific behavior with direct audio input

Fixed

Fix Qwen3-TTS not using audio from opt_narrator input
Fix Qwen3-TTS falling back to default narrator when audio-only is connected to opt_narrator

[4.19.8] - 2026-01-28

Changed

Improve model lifecycle management for VoiceDesign↔Base switching

Fixed

Fix Qwen3-TTS crashes when switching between Voice Designer and TTS Text
Fix device mismatch errors (CPU/CUDA tensor conflicts)
Fix language dropdown selection being overridden by English
Better handling of cached processors with correct engine references

[4.19.7] - 2026-01-28

Added

Narrator segments now correctly use the selected language

Fixed

Fix Qwen3-TTS language selection being ignored
Fix language dropdown selection not applying to generation
Resolves issue where Spanish/other languages switched to English

[4.19.6] - 2026-01-28

Changed

Improve handling of different voice reference formats (file paths, tensors, audio dicts)

Fixed

Fix Qwen3-TTS voice cloning error with tensor references
Fix 'Boolean value of Tensor with more than one value is ambiguous' error when using voice references

[4.19.5] - 2026-01-28

Added

Add CosyVoice model variant setting (RL/standard) for tags in settings menu
Prevent unnecessary model reload when TTS and VC use the same variant

Changed

Improve device management to properly reload CosyVoice components after unload

Fixed

Fix CosyVoice voice restoration device errors and add variant setting
Fix "Expected all tensors to be on the same device" errors during CosyVoice tag processing

[4.19.4] - 2026-01-28

Changed

Improve audio tensor handling across all engines

Fixed

Fix F5-TTS restore tag voice conversion failures
Fix voice conversion crashes when using restore tags with F5-TTS
Better compatibility between different TTS engines and voice conversion

[4.19.3] - 2026-01-28

Added

CosyVoice VC now reuses existing TTS engine instance when available
Prevents model thrashing when using CosyVoice for both TTS and voice restoration
Significantly faster VC processing when TTS model already loaded

Fixed

Fix CosyVoice voice conversion causing model reload errors
Eliminates device mismatch errors (tensors on different devices)

[4.19.2] - 2026-01-28

Added

Add user-selectable voice conversion engine for tags
New ComfyUI settings menu: Voice Restoration ( Tags)
Choose between ChatterBox 23-Lang (default), ChatterBox, or CosyVoice for voice restoration
Settings persist across ComfyUI restarts automatically
CosyVoice option enables native CosyVoice3 VC for seamless workflows
All engines support the same tag syntax with iteration control

[4.19.1] - 2026-01-28

Added

Add CosyVoice3 voice restoration support via tags
CosyVoice now supports tags for voice restoration using ChatterBox VC
Native CosyVoice paralinguistic tags (, , , etc.) remain fully functional
Step Audio EditX emotion/style/speed tags are blocked to avoid conflicts with CosyVoice's native system
Voice restoration applies after TTS generation as post-processing step

[4.19.0] - 2026-01-28

Added

Add Qwen3-TTS engine with voice design capabilities
New Qwen3-TTS engine supporting 0.6B and 1.7B models
Three model types: CustomVoice (9 preset speakers), VoiceDesign (create voices from text descriptions), Base (zero-shot voice cloning)
Support for 10 languages (Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian)
Voice Designer node creates custom voices from natural language descriptions
Full SRT subtitle support with character switching and all timing modes
Sage attention support for improved VRAM efficiency
Step Audio EditX inline edit tags for post-processing (emotions, styles, paralinguistic effects)
Auto-download system for all 6 model variants

Changed

Smart caching prevents duplicate voice generation

[4.18.0] - 2026-01-26

Added

Add VibeVoice Hindi community finetunes
Add vibevoice-hindi-1.5B support - Hindi-optimized 2.7B model
Add vibevoice-hindi-7B support - Hindi-optimized 9B model
Both models support Hindi and English languages
Auto-download from tarun7r HuggingFace repositories
Show download progress bars during model downloads

[4.17.0] - 2026-01-25

Added

Add ChatterBox Official 23-Lang Vietnamese community finetune
Add Vietnamese (Viterbox) support - community finetune with 24 languages total
Vietnamese model auto-downloads from dolly-vn/viterbox on first use
Select Vietnamese (Viterbox) from model version dropdown
Support for custom community finetunes with different vocab sizes

Changed

Expanded Vietnamese tokenization for better quality

[4.16.14] - 2026-01-24

Added

Add support for Apple Silicon Homebrew path (/opt/homebrew/lib)
Add support for Intel Mac Homebrew path (/usr/local/lib)

Fixed

Fix macOS install script failing on Apple Silicon with Homebrew
Fix dependency check not detecting Homebrew-installed libsamplerate and portaudio
Address issue #228

[4.16.13] - 2026-01-21

Fixed

Fix Step Audio EditX initialization on embedded Python 3.12+
Add setuptools dependency for Python 3.12+ distutils compatibility
Fix Step Audio EditX failing to load on embedded Python installations
Resolve model registration failures in FunASR bundled code
Address 'NoneType' object is not callable error during engine initialization

[4.16.12] - 2026-01-20

Added

Step Audio EditX engine now initializes correctly

Fixed

Fix Step Audio EditX import errors preventing engine initialization
Fix ModuleNotFoundError: No module named 'config.prompts'
Resolve module name collision with other config.py files in project

[4.16.11] - 2026-01-16

Fixed

Fix F5-TTS vocab.txt loading errors
Fix file not found error when loading E2-TTS models from HuggingFace
Fix fallback model loading path resolution
Resolve doubled 'infer/infer/' path issue (issue #222)

[4.16.10] - 2026-01-10

Added

Add automatic precision detection based on GPU capability (bf16 for RTX 30xx+, fp16 for older GPUs)

Fixed

Fix Step Audio EditX model loading and GPU compatibility
Fix dtype parameter error preventing model initialization
Improve compatibility with older GPUs (SM 7.5 and below)

[4.16.9] - 2026-01-07

Added

Add clear tooltips explaining TensorRT (<1% speedup) vs vLLM (~2.5x speedup)

Changed

Improve debug logging for optimization loading

Fixed

Fix CosyVoice TensorRT and vLLM loading issues
Fix TensorRT engine configuration (was missing required inputs)
Fix optimization settings not being preserved during lazy loading

[4.16.8] - 2026-01-05

Added

Add IndexTTS-2 QwenEmotion debug logging
Add debug output to help diagnose model loading issues
Shows exact path being searched for emotion model
Helps identify path mismatches on affected systems

[4.16.7] - 2026-01-03

Added

Voices configured in extra_model_paths.yaml now properly discovered
Voice files from shared model directories now appear in dropdowns

Fixed

Fix voices folder not discovered from extra_model_paths.yaml
Fixes issue where only local ComfyUI voices folder was scanned

[4.16.6] - 2026-01-03

Added

Characters without .txt files now properly clone voice from audio only

Fixed

Fix CosyVoice3 character switching without text references
Fix characters with audio but no .txt file using wrong voice
Automatically use cross_lingual mode when text reference is missing

[4.16.5] - 2026-01-03

Fixed

Fix CosyVoice3 PyYAML 6.0+ compatibility (enhanced fix)
Apply PyYAML compatibility patch across all CosyVoice entry points
Fix 'Loader' object has no attribute max_depth error more reliably
Ensure patch is applied before any yaml imports in the loading chain

[4.16.4] - 2026-01-02

Fixed

Fix CosyVoice3 model loading with PyYAML 6.0+ (tentative)
Fix 'Loader' object has no attribute max_depth error in CosyVoice3
Improve compatibility with newer PyYAML versions (6.0+)
Resolve CosyVoice3 initialization failures

[4.16.3] - 2026-01-02

Added

Address issues with embed_tokens and affine layers staying on CPU

Fixed

Fix CosyVoice3 device mismatch errors (tentative)
Fix 'Expected all tensors to be on the same device' error in CosyVoice3
Improve compatibility with PyTorch 2.9.1+cu130

[4.16.2] - 2026-01-02

Changed

Improve detection accuracy by checking actual module location instead of path keywords

Fixed

Fix IndexTTS-2 false positive detection in conda environments
Fix bundled IndexTTS-2 incorrectly flagged as external installation
Resolve compatibility issue when ComfyUI installed in directories with 'conda' in path

[4.16.1] - 2025-12-29

Fixed

Fix IndexTTS-2 QwenEmotion text emotion model failing to load
Fix error preventing QwenEmotion model from loading on some systems
Addresses compatibility issue with certain Python/modelscope versions
Note: Tentative fix - user testing required to confirm resolution

[4.16.0] - 2025-12-29

Added

Add CosyVoice3 TTS engine with multilingual support and voice conversion
New CosyVoice3 engine with 4 core languages (Chinese, English, Japanese, Korean) plus 5 additional languages
Voice cloning with zero-shot, instruction, and cross-lingual modes
Voice Changer (VC) with iterative refinement for quality improvement
Model variants: standard and RL-enhanced (improved quality, set as default)
Paralinguistic tag support for natural speech effects (laughter, breath, cough, etc.)
Character voice switching with [CharacterName] syntax
Language switching with [lang:code] syntax
SRT subtitle timing support for synchronized audio generation
Per-segment parameter control ([seed:42], [speed:1.5])
Live generation progress with token-by-token updates and ETA
Credit: Initial CosyVoice3 implementation by @tazztone (https://github.com/tazztone/TTS-Audio-Suite)

Fixed

Fix ChatterBox and RVC model discovery to work with custom model paths
Fix RVC audio chunking errors with short segments

[4.15.17] - 2025-12-28

Added

Parameters are now properly applied to generation
Add warning when using inline edit tags with multiple speakers in Native Multi-Speaker mode (edit tags only work with single narrator or Custom Character Switching mode)

Fixed

Fix VibeVoice inline parameters not applying correctly
Fix inline parameters like [seed:2|cfg:7] hitting cache incorrectly

[4.15.16] - 2025-12-28

Added

Russian model now downloads correctly from HuggingFace

Fixed

Fix ChatterBox Russian model download failing with 404 error
Resolves issue #210 crash on Russian text generation

[4.15.15] - 2025-12-28

Added

IndexTTS-2 SRT generation now works correctly with smart_natural

Fixed

Fix IndexTTS-2 SRT crash in smart_natural timing mode
timing mode - fixes crash that occurred during audio assembly

[4.15.14] - 2025-12-28

Added

that would cause installation errors

Fixed

Fix pyproject.toml configuration file syntax
ComfyUI Registry compatibility fix - removed invalid dependencies field

[4.15.13] - 2025-12-20

Added

Add automatic fallback to bundled vocab if download fails
Help users diagnose download issues more easily

Fixed

Improve F5-TTS model download error detection
Fix confusing errors when F5-TTS model downloads are incomplete
Provide clearer error messages indicating which files are missing

[4.15.12] - 2025-12-19

Changed

Improve Audio Editor standalone mode reliability

Fixed

Fix Step Audio EditX Audio Editor node errors
Fix Audio Editor failing with 'object has no attribute edit_single' error
Fix Windows compatibility issue with audio tokenization (torchcodec BytesIO)
Fix model not found error when using extra_model_paths.yaml configuration

[4.15.11] - 2025-12-18

Added

Removes GPU memory copy and node assignment warnings
These warnings are expected and don't affect functionality

Changed

Suppress harmless ONNX Runtime performance warnings in Step Audio EditX
Cleaner console output during Step Audio EditX generation

[4.15.10] - 2025-12-18

Added

Character voices discovery now logs only once per workflow
Eliminates duplicate 'Found X characters' messages in console

Changed

Character changes on refresh still show update notifications

Fixed

Fix character discovery log spam when using multiple character nodes

[4.15.9] - 2025-12-18

Added

Silent/gibberish audio generation
Generation ignoring max_new_tokens parameter
Model behaving unpredictably across different transformers versions
Implementation:
Detects weight tying via data pointer comparison
Detailed logging for future removal if stepfun-ai fixes their model

Fixed

Fix Step Audio EditX silence and max_new_tokens generation bug
Properly resolves transformers 4.54+ weight tying issue that caused:
Restores correct lm_head weights from model checkpoint
Comprehensive error handling with fallback behavior

[4.15.8] - 2025-12-17

Added

Detect conflicting utils.py files from other custom nodes
Add warning message showing which node is causing conflict

Fixed

Address issue #191: utils namespace conflict from other custom nodes
This is a tentative fix for users experiencing 'utils' is not a package errors.
Should resolve all engine loading failures
Note: This fix has been tested with simulated conflicts but not yet confirmed on affected user systems.

Removed

Automatically remove conflicts before TTS Suite loads

[4.15.7] - 2025-12-17

Added

Add GPU acceleration support for Step Audio EditX ONNX operations
New features:
Automatic ONNX GPU acceleration for CosyVoice vocoder and FunASR processing
Verified working on Windows with Python 3.10+
Improvements:
Clarified Step Audio EditX model directory structure in documentation

Fixed

Smart CUDA detection with automatic CPU fallback for compatibility
Better error handling for missing GPU libraries

[4.15.6] - 2025-12-17

Fixed

Fix Step Audio EditX silence/humming output issue
Resolves transformers 4.54+ compatibility by ensuring lm_head and embed_tokens weights are not incorrectly tied during model loading. Users will now get proper audio generation from Step Audio EditX without needing to re-download models or update config.json manually.

[4.15.5] - 2025-12-16

Added

Prevent AttributeError when Step Audio EditX model selection is None or invalid. Now gracefully defaults to auto-download instead of crashing.

Fixed

Fix Step Audio EditX model loading crash on Windows

[4.15.4] - 2025-12-16

Added

Add comprehensive validation for CosyVoice model files
Clear instructions for fixing corrupted model downloads

Fixed

Improve Step Audio EditX model validation and error handling
Better error messages when model files are missing or incomplete
Improved compatibility with different hyperpyyaml versions

[4.15.3] - 2025-12-12

Added

Add Character Voices node refresh support
Character Voices dropdown now refreshes with R key (no restart needed)
Character switching tags refresh automatically when pressing R
New voice files appear immediately after refresh
Console shows character count updates when changes detected

[4.15.2] - 2025-12-12

Added

Extension now immune to Python import path conflicts

Fixed

Fix extension failing to load when other custom nodes have utils.py files
Fix 'utils' is not a package error affecting all engines
Improve compatibility with other ComfyUI extensions

[4.15.1] - 2025-12-12

Added

Add IndexTTS-2 Low VRAM mode for 8GB VRAM GPUs
Add Low VRAM option to IndexTTS-2 Engine node
Keep models on CPU and move to GPU only when needed during generation
Prevents out-of-memory errors on 8GB VRAM cards
Successfully tested with <8GB VRAM usage throughout generation
Allows longer reference audio (tested up to ~2 minutes)
Trade-off: Slightly slower generation due to model swapping
Credit: Feature contributed by @CrispyToken in PR #186

[4.15.0] - 2025-12-12

Added

Step Audio EditX TTS Engine - AI-powered text-to-speech with zero-shot voice cloning
- Clone any voice from just 3-10 seconds of audio
- Natural-sounding speech with emotional control
- Memory-efficient int4/int8 quantization options (uses less VRAM)
- Character switching and per-segment parameter support
Step Audio EditX Audio Editor - Transform any TTS engine's output with AI
- 14 emotions: happy, sad, angry, surprised, fearful, disgusted, contempt, neutral, etc.
- 32 speaking styles: whisper, serious, child, elderly, neutral, and more
- Speed control: make speech faster or slower
- 10 paralinguistic effects: laughter, breathing, sigh, gasp, crying, sniff, cough, yawn, scream, moan
- Audio cleanup: denoise and voice activity detection
- Works with audio from ANY TTS engine (ChatterBox, F5-TTS, Higgs Audio, VibeVoice)
Universal Inline Edit Tags - Add audio effects directly in your text
- Easy syntax: "Hello this is amazing!"
- Works across all TTS engines using Step Audio EditX post-processing
- Tag types: <emotion>, <style>, <speed>, and paralinguistic effects
- Control intensity: <Laughter:2> for stronger effect, <Laughter:3> for maximum
- Voice restoration: <restore> tag to return to original voice after edits
Multiline TTS Tag Editor enhancements
- New tabbed interface for inline edit tag controls
- Quick-insert buttons for emotions, styles, and effects
- Better copy/paste compatibility with ComfyUI v0.3.75+
- Improved syntax highlighting and text formatting
New Example Workflows
- "Step Audio EditX Integration" - Basic TTS usage examples
- "Step Audio EditX - Audio Editor + Inline Edit Tags" - Advanced editing demonstrations
- Updated "Voice Cleaning" workflow with Step Audio EditX denoise option

Changed

Better memory management and model caching across all engines

Fixed

Step Audio EditX text being incorrectly combined when using newlines (now respects line breaks for natural pacing)
Step Audio EditX chunk combination method setting being ignored
Multiline TTS Tag Editor paste behavior in ComfyUI v0.3.75+

[4.14.36] - 2025-12-02

Added

Add chunking support to RVC voice conversion
RVC engine now supports intelligent chunking for long audio files
Default 30s chunks prevent OOM on long RVC conversions
All refinement passes apply correctly to each chunk
Seamless audio quality with crossfade blending

Changed

Smart and fixed chunking modes work with RVC conversions

Fixed

Full compatibility with existing RVC parameters and models

[4.14.35] - 2025-12-02

Added

Add intelligent chunking to Voice Changer node
Prevent OOM errors when converting long audio files (5+ minutes)
Fixed mode splits at exact intervals for continuous audio
Set max_chunk_duration to 0 to disable chunking entirely
Default 30s chunks work well for ChatterBox engines
RVC users can disable chunking or use larger chunk sizes
Full cache support for faster subsequent conversions

Changed

Smart mode splits at natural silences for seamless quality

Fixed

Fixes issue #184 (ChatterBox VC OOM on long audio)
Implements feature request from issue #24 (chunking support)

[4.14.34] - 2025-12-02

Added

Models now correctly detect if they're already in the right subdirectory

Fixed

Fix RVC HuBERT models creating nested subdirectories
Fix Chinese HuBERT and other transformers models creating double-nested folders

[4.14.33] - 2025-12-02

Changed

Cleaner console output during generation

Removed

Remove IndexTTS-2 debug output from console
Remove debug messages showing audio data types and ranges

[4.14.32] - 2025-12-02

Added

Add support for language-specific HuBERT models (Japanese, Korean, Chinese, Large)
Add automatic model organization into subdirectories

Fixed

Fix RVC voice conversion HuBERT model loading issues
Fix Japanese HuBERT download failures (401 errors)
Fix user HuBERT model selection being ignored
Fix config.json conflicts between different HuBERT models
Add helpful error messages when HuBERT/RVC model dimensions don't match
Fix compatibility with transformers 4.46+ and Python 3.13

[4.14.31] - 2025-11-20

Changed

Improve audio processing robustness across all engines

Fixed

Fix audio dimension errors when using Wave Analyzer output
Fix Wave Analyzer segmented audio causing tensor dimension errors in TTS generation
Fix voice conversion nodes failing with unexpected audio tensor dimensions

[4.14.30] - 2025-11-20

Added

Move Voice Fixer to Audio Processing category alongside Noise/Vocal Removal
Much faster model downloads via HuggingFace instead of Zenodo
Show live download progress in console

Fixed

Fix Voice Fixer model loading and downloads
Fix Voice Fixer failing to find downloaded models (vocoder path bug)
Fix Windows Unicode errors during downloads

[4.14.29] - 2025-11-19

Added

Add informative warnings for known PyTorch 2.9.0+xpu beam sampling issues

Changed

Improve cache management for Intel GPU devices

Fixed

Add IndexTTS-2 XPU (Intel GPU) compatibility improvements
Fix CPU core pinning at 100% on PyTorch 2.8.0+xpu during audio processing
Fix kaldi.fbank failures on PyTorch 2.8.0+xpu and 2.9.0+xpu
Fix tensors incorrectly moving to XPU when CPU device is explicitly selected
Note: Changes address GitHub issue #167 but are untested due to lack of Intel GPU hardware

[4.14.28] - 2025-11-19

Added

Italian language prefix now works with any capitalization ([it], [IT], [It], etc.)
Prevents accidental double-prefixing when users add manual language markers

Fixed

Fix Italian ChatterBox language prefix detection

[4.14.27] - 2025-11-15

Added

Streaming now works correctly with batch_size > 1 for parallel processing
All streaming segments complete successfully without NoneType errors

Fixed

Fix ChatterBox streaming batch processing
Fix tuple unpacking errors in multi-character streaming mode
Fix transformers compatibility issue causing streaming segment crashes
Add better error logging for debugging streaming failures

[4.14.26] - 2025-11-15

Added

Eliminates 'Model not wrapped - using legacy path' warning
Ensures Clear VRAM button works correctly for all cached model instances

Fixed

Fix model wrapper consistency for Clear VRAM support
Fix cached models returning unwrapped instances instead of ComfyUI wrappers

Removed

Remove obsolete model unwrapping code from streaming adapter

[4.14.25] - 2025-11-15

Added

Add automatic 50000x scaling to make exaggeration parameter more intuitive (use normal 0.5-2.0 values instead of 10000+)
Expand exaggeration range to 0.0-5.0 for experimentation

Changed

Improve ChatterBox Official 23-Lang exaggeration control
Update tooltip with honest warning that this parameter has minimal effect in multilingual models due to ResembleAI training limitations

Fixed

Note: Classic ChatterBox exaggeration works as documented; issue only affects multilingual v1/v2 models

[4.14.24] - 2025-11-15

Fixed

Fix ChatterBox Official 23-Lang audio generation failure
Fix crash when generating long audio with chunking enabled in single-character mode
Resolves tuple unpacking error (expected 3 values, got 4)

[4.14.23] - 2025-11-14

Added

Standardize all engine model loading through unified interface
Add centralized engine capability registry
Add generic fallback handler for language switching
Standardize model configuration across all engines
Clean up legacy wrapper-searching code patterns

Changed

Improve architecture:
Better ComfyUI integration for memory management

Fixed

Complete model loading refactor addressing issue #14
Fix model loading architecture issues:
Fix 'Clear VRAM' button not working with TTS models
Fix Higgs Audio models loading twice (cache duplication bug)
Improve error handling consistency across all TTS engines

Removed

Remove duplicate SmartModelLoader caching system
Remove 'Wrapper not found' warning messages

[4.14.22] - 2025-11-13

Added

Device Management:
Cache 'auto' device resolution for consistent cache keys across sessions
Consolidate device resolution to use single torch_device_resolver function
Prevent unnecessary model reloads when switching device settings
Architecture:

Changed

Performance & VRAM:
Improve cache hit rates by resolving devices before cache key generation
Simplify model loading to unified interface for better maintainability

Fixed

Fix model loading cache issues and improve device handling
Fix device type mismatches in model loading comparisons
Fix Python 3.12 VRAM spikes in voice conversion operations

Removed

Remove redundant SmartModelLoader caching system from F5-TTS

[4.14.21] - 2025-11-13

Added

Voice conversion now works even if perth library lacks PerthImplicitWatermarker

Fixed

Fix ChatterBox Voice Conversion initialization crash
Resolve watermarker attribute errors in voice conversion models
Watermarking remains disabled by default for maximum compatibility

[4.14.20] - 2025-11-08

Added

Preview video now generates and displays with mouth movement markers
Annotated frames show movement detection with confidence visualization

Fixed

Fix Silent Speech Analyzer video preview not displaying

[4.14.19] - 2025-11-07

Added

Add Intel XPU (Intel GPU) support with automatic detection
Tentative XPU support for Intel GPU acceleration
Auto-detection in device selection (MPS > CUDA > XPU > CPU)
Available on all TTS and processing nodes
Requires separate Intel PyTorch extension installation
Note: XPU support not validated on Intel hardware yet - feedback welcome

[4.14.18] - 2025-11-07

Added

when processing audio. Root cause: local numpy imports inside exception
handlers shadowed the module-level import, leaving 'np' undefined when
exceptions didn't occur but code later tried to use it.
Fixed by removing redundant local imports and using module-level numpy.

Fixed

Fix Higgs Audio numpy scope error
Higgs Audio engine would crash with 'cannot access local variable np'

[4.14.17] - 2025-11-07

Added

Root issues fixed:
PyTorch 2.9 changed torchaudio.load() to return raw int16 values instead of normalized audio in [-1, 1] range
TorchCodec DLL incompatibility on Windows prevented voice conversion and other features from working
Character Voices node wasn't normalizing audio from files, causing 1.5x amplification
Changes made:
Centralized audio loading with automatic int16 normalization across the suite
Windows voice conversion and F5-TTS now working without TorchCodec DLL errors
Comprehensive documentation of fixes in code comments

Fixed

Fix IndexTTS-2 audio distortion on PyTorch 2.9 with Windows voice conversion compatibility
Issue resolved: Audio distortion in IndexTTS-2 and multiple engines (F5-TTS, ChatterBox, Voice Changer) on PyTorch 2.9
Global PyTorch 2.9 compatibility patch applied at startup (before any module imports torchaudio)
Safe audio loading utilities with scipy fallback for better compatibility

[4.14.16] - 2025-11-06

Added

RVC and Hubert models now properly registered with ComfyUI
Pitch extraction models automatically cleaned up after conversion
Memory is intelligently freed when VRAM is needed by other operations
ComfyUI's Clear VRAM button now properly unloads all RVC models

Fixed

Fix RVC voice conversion VRAM buildup

[4.14.15] - 2025-11-06

Added

Added amplitude normalization for IndexTTS-2 generated audio
Audio should now have normal volume levels consistent with Windows behavior

Fixed

Fix IndexTTS-2 generating extremely loud and distorted audio on Linux
Tentative fix for Linux-specific audio distortion issue

[4.14.14] - 2025-11-06

Changed

Improve multilingual model caching to recognize same model across different language parameters

Fixed

Fix ChatterBox Official 23-Lang SRT processing and model caching issues
Fix batch_size=0 not disabling streaming in SRT mode (was always using streaming)
Fix model reloading when switching between SRT and TEXT nodes with same engine on first execution

[4.14.13] - 2025-11-06

Added

Prevent VRAM accumulation when switching between different RVC models
Properly unload old models when loading new ones to free GPU memory
Support multiple RVC and Hubert model variants without VRAM leaks

Changed

Preserve model caching performance by only unloading when model actually changes

Fixed

Fix RVC VRAM buildup when switching between character voice models
Fix chimpmunk audio output caused by device dtype mismatches

[4.14.12] - 2025-11-05

Added

Clear cached pitch extraction models (RMVPE, FCPE) after each conversion completes
Prevent OutOfMemory errors when performing multiple conversions in succession

Changed

Maintain model caching performance for frequently-used RVC and Hubert models

Fixed

Fix RVC progressive VRAM buildup on repeated voice conversions
Resolve VRAM accumulation that occurred with each RVC conversion

[4.14.11] - 2025-11-05

Added

Eliminated torchcodec requirement that needed FFmpeg system installation
All audio loading now uses automatic fallback mechanisms
Support for WAV, MP3, FLAC and other formats maintained
Installation no longer requires external FFmpeg setup

Fixed

Remove FFmpeg system dependency for better Windows compatibility
Better error handling with graceful degradation

[4.14.10] - 2025-11-05

Added

Character switching now works correctly with proper character voice files
Timing reports show proper emoji indicators (🐰 compress, 🐌 expand) matching other engines
Unified timing engine functions ensure consistent report format across all modes

Fixed

Fix ChatterBox Official 23-Lang SRT character switching and timing report consistency
Audio segments maintain correct SRT order despite language-grouped processing
Fixes issue #157

[4.14.9] - 2025-11-05

Added

All TTS engines continue to work with compatible package versions

Fixed

Fix package installation dependency conflicts
Fix numpy being downgraded before version protection runs
Fix pydantic/pydantic-core downgrades from gradio
Fix rotary-embedding-torch regression during audio-separator install
Improve installation sequence for proper dependency resolution

[4.14.8] - 2025-11-05

Added

Add comprehensive MPS device support across all TTS engines
All TTS engines now properly support Apple Silicon MPS when 'auto' device is selected
Fixed 'auto' device defaulting to CPU on Apple Silicon Macs
Intelligent device detection: MPS (Apple Silicon) > CUDA (NVIDIA) > CPU
RVC voice conversion now works with MPS acceleration

Changed

Improved performance on M-series Macs with proper GPU utilization

[4.14.7] - 2025-11-05

Added

Fixed FileNotFoundError when using RVC with lower-end NVIDIA GPUs
RVC now properly locates configuration files in all installation paths
Resolves crashes during voice conversion on GTX 1080/1080 Ti hardware

Fixed

Fix RVC voice conversion on GTX 1080/1080 Ti GPUs

[4.14.6] - 2025-11-04

Fixed

Fix audio loading and German model download issues
Fixed missing torchcodec dependency preventing audio file loading in Character Voices node
Fixed RVC voice conversion dependency checking for proper detection
Fixed German (havok2) ChatterBox model download 404 error by using correct model folder

[4.14.5] - 2025-11-04

Added

Add MPS (Apple Silicon) GPU acceleration support
Automatic device detection prioritizes available hardware
'auto' mode now detects: Apple Silicon MPS → NVIDIA CUDA → CPU fallback
MPS device option added to all engine configuration nodes

Changed

All TTS engines now support Apple Metal Performance Shaders (MPS)

Fixed

Full backward compatibility - no breaking changes

[4.14.4] - 2025-11-04

Added

Models now load from TTS/voicefixer/ directory instead of ~/.cache/
Respects extra_model_paths.yaml for shared model directories

Changed

Better integration with ComfyUI model management system

Fixed

Fix Voice Fixer model loading to use correct directory

[4.14.3] - 2025-11-03

Added

Undo/redo history now preserved across page reloads
Multiple tag editors in same workflow maintain separate data
Text persists correctly when reloading workflows
History resets appropriately when opening different workflow files
Fixed character truncation in localStorage during saves

Fixed

Fix Multiline TTS Tag Editor history and workflow persistence

[4.14.2] - 2025-11-03

Added

Text now persists when saving and loading workflows in different ComfyUI installations
Undo/redo history is properly preserved when reloading within same session
Local edits persist across reloads while maintaining workflow-shared values

Fixed

Fix Multiline TTS Tag Editor text persistence in saved workflows

[4.14.1] - 2025-11-01

Added

Regular typing now works without triggering ComfyUI shortcuts
Alt+L/C/P/Z and Alt+1/2/3 shortcuts now respond correctly
Enter key creates newlines without delays or caret jumping
Improved overall text editing experience

Fixed

Fix keyboard shortcuts and text input in tag editor

[4.14.0] - 2025-10-31

Added

Add 🏷️ Multiline TTS Tag Editor with rich text editing and tag management
Major Features:
Interactive tag editor with character/language/parameter dropdowns
Preset system to save and load up to 3 configurations
Full undo/redo history with keyboard shortcuts
Syntax highlighting for tags and SRT timings
Resizable fonts (2-120px) and 16 web-safe fonts
Customizable UI scaling and sidebar width
Per-node unique storage for multiple editors in same workflow
Toast notifications for feedback
Tag formatting and validation

Changed

Select text to prepend tags or modify existing tags

[4.13.2] - 2025-10-30

Added

Text-based emotion analysis now works with current transformers versions

Fixed

Fix IndexTTS-2 text emotion support with modern transformers (#90)
Resolve QwenEmotion model loading error with transformers 4.50+
Improved compatibility with various transformers library versions

[4.13.1] - 2025-10-29

Fixed

Fix ChatterBox watermarking library import issue
Resolve ModuleNotFoundError when perth is not installed in ComfyUI Manager
Watermarking now gracefully handles missing optional dependency
Improves compatibility with various ComfyUI installation methods

[4.13.0] - 2025-10-29

Added

New Voice Fixer node for professional audio restoration (removes noise, reverberation, clipping)
Support 3 restoration modes for different audio degradation levels

Changed

Add Voice Fixer audio restoration and improve audio processing
Optimize dependencies by replacing torchlibrosa with librosa integration
Improve Wave Analyzer to preserve stereo audio quality throughout pipeline
Better organized model selection dropdown for vocal removal (grouped by category)

Fixed

Fix audio playback not updating when input changes

[4.12.1] - 2025-10-29

Added

Higgs Audio now uses consistent fixed seed across segments
Per-segment seed control available via segment parameters

Changed

Better cache consistency when editing or reordering subtitles

Fixed

Fix Higgs Audio SRT seed behavior

[4.12.0] - 2025-10-29

Added

Add per-segment parameter switching across all TTS engines
Override generation parameters (seed, temperature, CFG, speed, etc.) per segment
Use inline tag syntax: [Alice|seed:42|temperature:0.5] for fine-grained control
Works with all engines: ChatterBox, F5-TTS, Higgs Audio 2, VibeVoice, IndexTTS-2
Apply parameters to individual SRT subtitle lines
Order-independent parameter specification
Engine-specific parameters automatically filtered
Backward compatible with character and language switching
Per-segment parameters override node defaults without affecting other segments

[4.11.26] - 2025-10-28

Added

Add interrupt handling for all TTS engines
All TTS engines now support operation cancellation:
Users can stop SRT generation at any point
Interruption checks at segment processing boundaries
Character switching and chunking operations are interruptible
Added documentation for future engine implementations

Fixed

Graceful error handling with informative messages

[4.11.25] - 2025-10-27

Changed

Improve memory management with proper CUDA/CPU device handling across all engines

Fixed

Fix Clear VRAM button for all TTS engines
Fix ChatterBox Classic, ChatterBox 23-Lang, F5-TTS, Higgs Audio, VibeVoice, RVC, and Index-TTS models not unloading from GPU when Clear VRAM is clicked
Fix device mismatch errors after unloading and reloading TTS models
Fix VRAM spike issues when reloading RVC and VibeVoice models

[4.11.24] - 2025-10-27

Added

Ensure model components properly reload to GPU after VRAM optimization

Fixed

Fix ChatterBox Official 23-Lang device placement causing crashes after unloading
Fix crashes on second generation after clicking 'Clear VRAM' button
Fix device mismatch errors when models are moved between CPU and GPU
Fix 'Clear VRAM' button not working correctly after model reload

[4.11.23] - 2025-10-26

Fixed

Fix PyTorch installation failing on Windows Python 3.13
Fix invalid pip requirement syntax preventing installation on all platforms
Add graceful fallback if RVC dependencies unavailable (build tools missing)

[4.11.22] - 2025-10-25

Added

Complete cleanup of deprecated transformers API usage

Fixed

Fix remaining torch_dtype deprecation warning in Higgs Audio

Removed

Remove final torch_dtype deprecation warning during model initialization

[4.11.21] - 2025-10-25

Fixed

Fix Higgs Audio compatibility with transformers 4.57.1+
Fix audio generation failing with transformers 4.57.1 and newer
Maintain full backward compatibility with older transformers versions

Removed

Remove deprecated torch_dtype warnings during model loading

[4.11.20] - 2025-10-23

Added

Clearer diagnostic messages during startup
Add [TTS Audio Suite] prefix to character voice discovery messages
Makes it easier to identify which extension is reporting information

[4.11.19] - 2025-10-23

Added

Faster ComfyUI startup with lazy engine loading
Reduce blocking operations during ComfyUI initialization
Cache voice character discovery for faster repeated loading

Changed

Performance improvements:
Optimize all 7 TTS engine nodes to load instantly on startup

Fixed

Move dependency validation to background thread

[4.11.18] - 2025-10-23

Added

Removed verbose 'available' and 'loaded' success messages during initialization
Keep important diagnostic information for troubleshooting
Improved code clarity by removing complex message filtering attempts

Changed

Cleaner startup output with reduced logging spam
Better user experience with less console noise on startup

[4.11.17] - 2025-10-23

Added

Resolves issues when using rgthree Any Switch node

Changed

Reduces console spam with smart model download detection
Improves user experience with cleaner logs

Fixed

Fix IndexTTS-2 and other engines failing on first run after reboot
Fix engine parameters corrupted by ComfyUI JSON serialization

[4.11.16] - 2025-10-22

Changed

Improve model reuse reliability after VRAM management operations

Fixed

Fix ChatterBox CUDA device mismatch error on subsequent generation runs
Resolve 'tensors on different devices' error that occurs on second+ TTS generation with cached models
Fix tokens not being moved to GPU when models are reloaded after CPU offloading

[4.11.15] - 2025-10-22

Changed

Improve stability when switching between CPU and CUDA device placement

Fixed

Fix Higgs Audio cache errors in Memory Safe mode
Fix AttributeError when resetting DynamicCache in Memory Safe mode (CUDA Graphs OFF)
Fix cache not being recreated after model unload/reload cycle
Better error handling and logging for cache operations

[4.11.14] - 2025-10-22

Added

IndexTTS-2 uses fallback text processing on Windows (fully functional)
Add helpful note in installation summary explaining why

Changed

Improve Windows install experience for IndexTTS-2
User impact: Much cleaner install process on Windows with no confusing errors

Fixed

WeTextProcessing no longer shows confusing error messages on Windows install.
Windows users get clean installation without error spam

[4.11.13] - 2025-10-22

Fixed

Fix VibeVoice num_hidden_layers compatibility issue
Fix VibeVoice generation error: 'VibeVoiceConfig' object has no attribute 'num_hidden_layers'
Add compatibility patch for transformers 4.51.3+ DynamicCache initialization
Implements fix from wildminder/ComfyUI-VibeVoice v1.5.1 for FushionHub fork

[4.11.12] - 2025-10-21

Added

User impact: IndexTTS-2 now properly downloads and caches the CampPlus speaker embedding model

Fixed

Fix IndexTTS-2 CampPlus model download failure
Resolves issue #144 where IndexTTS-2 fails to load with FileNotFoundError.
CampPlus model now automatically downloads to the correct location.

[4.11.11] - 2025-10-20

Added

First generation works fine
Now ensures conditioning data moves with model to CUDA
User impact: ChatterBox Official 23-Lang now works correctly when running multiple generations

Fixed

Fix ChatterBox Official 23-Lang CUDA device mismatch on repeated generations
Resolves issue #143 where generation fails on second run after model detach.
After VRAM management moves model to CPU, second run fails with device mismatch

[4.11.10] - 2025-10-19

Added

Add optional SSL bypass for W2V-BERT model downloads (opt-in via TTS_DISABLE_SSL_VERIFY=1)
SSL verification is temporarily disabled only during model loading, then restored

Fixed

Fix IndexTTS-2 SSL certificate errors for users in restricted networks
Resolve certificate verification failures in China and corporate networks

[4.11.9] - 2025-10-14

Fixed

Fix VibeVoice not detecting custom models in extra_model_paths.yaml
Fix support for .safetensor file extension (some models use singular form)
Fix custom model paths not being searched correctly
Improve compatibility with user-organized model directories

[4.11.8] - 2025-10-14

Fixed

Fix F5-TTS engine failing to load in ComfyUI portable
Fix protobuf version mismatch preventing all TTS engines from loading
Make wandb dependency optional (only needed for training, not inference)
Resolves 'Detected incompatible Protobuf Gencode/Runtime versions' error in ComfyUI portable installations

[4.11.7] - 2025-10-13

Fixed

Fix RVC HuBERT model loading issues
Fix 'NoneType' object is not subscriptable error during voice conversion
Improve compatibility with chinese-hubert-base.pt models
Better error messages when model conversion fails

[4.11.6] - 2025-10-10

Added

All IndexTTS-2 dependencies now download to TTS folder as intended

Changed

Better organization of model files following project standards

Fixed

Fix IndexTTS-2 model download policy violations
Fix w2v-bert-2.0 model downloading to wrong directory
Fix MaskGCT semantic codec downloading to wrong directory
Fix BigVGAN vocoder downloading to wrong directory

[4.11.5] - 2025-10-10

Added

HuBERT now downloads to TTS/hubert/hubert-base-ls960/ following TTS structure
Consistent behavior with all other engines for model organization
Properly respects extra_model_paths.yaml for custom model directories
All engines now follow unified download policy: local > cache check > TTS download

Fixed

Fix Higgs Audio downloading dependencies to .cache instead of TTS folder
Fix Higgs Audio HuBERT model downloading to .cache against our unified policy

[4.11.4] - 2025-10-10

Added

Support multiple parent folder naming conventions (HiggsAudio/, higgs_audio/, etc.)
Allow flexible subfolder naming like other TTS engines
Models now properly detected from configured shared directories
Consistent behavior with ChatterBox, F5-TTS, VibeVoice, and IndexTTS-2

Fixed

Fix Higgs Audio models not being detected from custom directories
Fix Higgs Audio not respecting extra_model_paths.yaml configuration

[4.11.3] - 2025-10-10

Added

Support both TTS/vibevoice/ and TTS/VibeVoice/ folder naming
Allow flexible subfolder naming like other TTS engines
Models now properly detected from configured shared directories

Fixed

Fix VibeVoice models not being detected from custom directories
Fix VibeVoice not respecting extra_model_paths.yaml configuration

[4.11.2] - 2025-10-10

Added

Phoneme Text Normalizer now handles Polish ł correctly
Affects Unicode Decomposition method in experimental text preprocessing

Fixed

Fix Polish text normalization adding extra ł characters
Fix Polish text showing doubled ł characters (słynie -> slłynie)

[4.11.1] - 2025-10-08

Added

Add honest documentation about experimental nature of special tokens
Note: v2 special tokens exist in vocabulary but may produce minimal effects

Fixed

Fix ChatterBox v2 special token conversion
Fix v2 emotion tags not being converted from to [emotion] format
Fix incorrect token names in changelog and examples
Link to ResembleAI issue showing lack of official documentation and limited effectiveness

[4.11.0] - 2025-10-08

Added

Add ChatterBox Official 23-Lang v2 with emotion and sound effects
New v2 model adds 30+ special tokens for expressive speech:
Emotions: , , , ,
Sounds: , , , ,
Vocal effects: , ,
Features:
Select v1 or v2 model in ChatterBox engine configuration
Auto-download v2 model files when selected
Use angle brackets to avoid conflicts with character switching
Both v1 and v2 models can coexist in same installation
See docs/CHATTERBOX_V2_SPECIAL_TOKENS.md for complete token list and usage examples

Fixed

Full compatibility with existing character voices and pause tags
Bug fixes:
Fix v1 and v2 models properly isolated (separate caches)
Fix v2 model auto-download with correct file paths
Fix v1 model loading with correct vocabulary size

[4.10.16] - 2025-10-07

Added

Add chunking support for VibeVoice native multi-speaker mode
Support chunk combination method selection from unified node

Changed

Improve text chunking to respect sentence boundaries

Fixed

Fix VibeVoice local model loading and chunking
Fix local: prefix models failing with 'Unknown model' error

[4.10.15] - 2025-09-29

Changed

Better fallback support for offline tokenizer downloads

Fixed

Fix VibeVoice tokenizer download errors
Fix VibeVoice failing with 'download_huggingface_model' attribute error
Improve error handling when HuggingFace cache is disabled

[4.10.14] - 2025-09-27

Changed

Improve memory management to prevent unnecessary model reloading

Fixed

Fix VibeVoice performance issues with auto device mode
Fix models being moved to CPU after first generation
Resolve slow subsequent generations when using auto device setting
Fix VibeVoice auto device detection staying on GPU consistently

[4.10.13] - 2025-09-26

Changed

Improve cache accuracy for different VibeVoice configurations

Fixed

Fix VibeVoice cache not invalidating when changing attention modes
Fix cache incorrectly reusing audio when switching between attention modes

[4.10.12] - 2025-09-26

Added

Add accelerate version to startup diagnostics

Changed

Better support for different environment configurations

Fixed

Fix VibeVoice compatibility with accelerate library versions
Fix VibeVoice failing to load with older accelerate versions
Add smart compatibility patches for transformers + accelerate mismatches
Improve error handling for missing accelerate functions

[4.10.11] - 2025-09-26

Added

Add automatic fallback if SageAttention fails during generation
Credits: SageAttention approach adapted from @Enemyx-net's implementation

Changed

Improve generation speed and reliability with SageAttention enabled

Fixed

Fix VibeVoice SageAttention generating infinite noise and corrupting model state
Fix SageAttention generating 2+ minute audio with weird noise instead of proper speech
Fix model state corruption requiring ComfyUI restart when switching attention modes
Fix VibeVoice not stopping generation at proper endpoints

[4.10.10] - 2025-09-26

Changed

Improve IndexTTS-2 reliability and user experience

Fixed

Fix IndexTTS-2 import conflicts with external installations
Fix import errors when external IndexTTS packages are installed
Add clear error messages with solution steps for conflicting installations
Resolve 'No module named indextts.gpt.model_v2' errors reported by users

[4.10.9] - 2025-09-26

Fixed

Fix severe VRAM spikes on Python 3.12 with automatic compatibility patch
Fix multi-GB VRAM spikes occurring before and after TTS generation
Add automatic ComfyUI compatibility system (no user action required)
Fix affects all TTS engines: ChatterBox, F5-TTS, IndexTTS-2
Resolve ComfyUI v0.3.57+ regression in Python 3.12 environments
Patch only applies when needed, preserving performance on other Python versions

[4.10.8] - 2025-09-25

Added

Add warnings when virtual environment Python differs from detected version

Changed

Better guidance for Windows users on using 'python' vs 'py' commands

Fixed

Fix installer Python version detection on Windows
Fix install script showing wrong Python version when using 'py' command

Removed

Remove emojis from installer messages to prevent encoding errors

[4.10.7] - 2025-09-25

Added

Add support for [1] [2] [3] [4] numeric character tags in native multi-speaker mode
Support mixed usage like [1] [2] [Alice] mapping to Speakers 1, 2, 3

Changed

Improve voice generation quality with true zero-shot hybrid mode
Increase CFG scale maximum from 5.0 to 10.0 for better control

Fixed

Add VibeVoice compatibility improvements with wildminder format
Full backward compatibility with existing character tag systems

[4.10.6] - 2025-09-23

Added

No changes needed for existing setups - all voice discovery now works correctly

Changed

Improve path comparison logic to properly detect duplicate directories

Fixed

Fix voice discovery not finding voices in extra_model_paths.yaml directories
Fix voice dropdown not showing voices from shared directories configured in extra_model_paths.yaml
Fix character voice switching not finding voices in configured shared directories

[4.10.5] - 2025-09-23

Added

No functionality lost - features still work with appropriate fallbacks

Changed

Add FFmpeg availability check to startup logs for better troubleshooting
Improve vocal separation to automatically fallback to WAV when FFmpeg unavailable

Fixed

Add graceful FFmpeg dependency handling with fallbacks
Better error messages when FFmpeg missing instead of cryptic subprocess errors

[4.10.4] - 2025-09-23

Added

No changes required for existing users

Fixed

Fix VibeVoice generation compatibility with transformers 4.56+
Fix VibeVoice generation failing with newer transformers versions
Maintain compatibility with older transformers versions

[4.10.3] - 2025-09-23

Changed

Improve Japanese text handling in F5-TTS speech editing

Fixed

Fix F5-JP Japanese text processing issues
Fix F5-TTS Japanese model incorrectly applying Chinese pronunciation
Update F5-JP to use newer model checkpoint for better compatibility

[4.10.2] - 2025-09-23

Changed

Improve stability for voice reference processing

Fixed

Fix ChatterBox Official 23-Lang intermittent device mismatch errors
Fix random crashes during subsequent generations
Resolve GPU/CPU tensor compatibility issues

[4.10.1] - 2025-09-23

Changed

Add comprehensive tooltips to 📝 Phoneme Text Normalizer node for better user guidance
Update PROJECT_INDEX.md with new Text Processing Nodes section
Improve user experience with detailed parameter explanations

Fixed

Add missing F5-TTS workflow example and improve node documentation
Add missing workflow file that was referenced in README and GitHub issues

[4.10.0] - 2025-09-23

Added

Add universal multilingual text preprocessing for improved pronunciation
Add 📝 Phoneme Text Normalizer node for multilingual TTS preprocessing
Add IPA phonemization support with automatic language detection
Add Unicode decomposition and character mapping fallbacks
Add F5-Polish model auto-download for high-quality Polish synthesis
Compatible with all TTS engines (F5-TTS, ChatterBox, Higgs Audio, etc.)

Changed

Improve pronunciation quality across all supported languages

Fixed

Fix Polish, German, French special character pronunciation in F5-TTS

[4.9.25] - 2025-09-21

Added

Add clear warnings for users with outdated transformers versions
Provide exact upgrade commands when version conflicts detected

Fixed

Fix transformers compatibility issues
Remove problematic compatibility patches that masked upgrade requirements
Better error messages for DynamicCache property setter issues

[4.9.24] - 2025-09-21

Added

Unify fallback handling across Windows, Mac, and Linux systems
Users can safely uninstall incorrect 'tn' package if previously installed

Fixed

Fix incorrect text normalization dependency causing confusion
Remove wrong 'tn' package dependency (time utilities instead of text processing)
Fix IndexTTS-2 text normalization to only require WeTextProcessing or wetext
Better error messages when text processing packages unavailable

[4.9.23] - 2025-09-18

Added

Add automatic fallback to wetext package if WeTextProcessing unavailable
Implement graceful degradation to basic text processing as last resort
Ensure IndexTTS-2 always works even without advanced normalization

Fixed

Fix IndexTTS-2 graceful fallback for missing text normalization
Fix IndexTTS-2 crashes when WeTextProcessing fails to install
Better error messages showing which text processing is available

[4.9.22] - 2025-09-18

Added

Add automatic fallback to compatible wetext package
Ensure text normalization always works with graceful degradation

Fixed

Fix IndexTTS-2 text processing installation issues
Fix WeTextProcessing installation failures on Windows Python 3.13
Better error messaging for text processing capabilities

[4.9.21] - 2025-09-18

Changed

Better reliability for users with interrupted downloads or storage issues

Fixed

Fix IndexTTS-2 auto-download and handle partial/corrupted model files
Fix IndexTTS-2 models now automatically download on first use
Fix partial downloads are detected and completed automatically
Fix corrupted or incomplete model files trigger smart re-download
Improve error messages show exactly which files are missing

[4.9.20] - 2025-09-18

Added

Based on updates from https://github.com/index-tts/index-tts (September 8-18, 2025):

Changed

Update IndexTTS-2 with latest upstream improvements from official repository
Improve emotion vector logging shows normalization and alpha scaling clearly
Better memory management eliminates VRAM accumulation during multiple generations

Fixed

Fix emotion vectors now auto-normalize to prevent voice identity loss (commit 8aa8064)
Fix memory issues with automatic audio truncation to 15 seconds (commit 0828dcb)
Fix persistent cache buildup during batch processing (commit 64cb31a)
Fix CUDA performance via corrected BigVGAN kernel imports (commits ee23371, e409c4a)

[4.9.19] - 2025-09-18

Added

Use proper PyPI package name for Chinese/English text normalization
Maintain tn fallback for systems where WeTextProcessing fails

Fixed

Fix IndexTTS-2 text normalization dependency package name
Fix incorrect wetext package reference to correct WeTextProcessing
Improve IndexTTS-2 text processing compatibility across platforms

[4.9.18] - 2025-09-18

Added

Add cumulative click counters showing total adjustments (+0.1, +0.2, etc.)

Changed

Improve IndexTTS-2 emotion radar dragging and click feedback
Better visual feedback for both reducing all emotions and increasing individual ones
Improve button text positioning in radar chart interface

Fixed

Fix emotions randomly switching when dragging near center point
Fix emotions bouncing back when dragged past zero to opposite side

[4.9.17] - 2025-09-17

Added

IndexTTS-2 now respects extra_model_paths.yaml configuration
Downloads go to configured shared directories instead of ComfyUI folder
Model discovery works across all configured paths
Dropdown shows local: prefix for found models, model name for download

Fixed

Fix IndexTTS-2 extra_model_paths support

[4.9.16] - 2025-09-17

Added

Ensure consistent GPU usage when CUDA is available

Changed

Better performance for multi-generation workflows

Fixed

Fix VibeVoice auto device mode performance
Fix auto mode defaulting to CPU for subsequent generations

[4.9.15] - 2025-09-17

Added

Ensure both primary and fallback text normalizers are available

Fixed

Fix IndexTTS-2 missing text normalization dependency on Linux
Add missing tn package dependency for IndexTTS-2 text processing
Fix IndexTTS-2 failing with tn.chinese module not found error
Improve compatibility with Linux systems where wetext fails to install

[4.9.14] - 2025-09-17

Added

Provide specific guidance when audio files are too long
Recommend optimal audio length (under 30 seconds)
Show clear solutions when generation fails due to memory limits

Fixed

Fix IndexTTS-2 crashes with large audio files
Add graceful out-of-memory error handling

[4.9.13] - 2025-09-17

Added

Ensure stable module loading for IndexTTS-2 initialization

Changed

Improve bundled module import handling to prevent conflicts

Fixed

Fix IndexTTS-2 KeyError import failure with bundled modules
Fix IndexTTS-2 failing with KeyError indextts during module resolution
Better compatibility with different Python import scenarios

[4.9.12] - 2025-09-17

Added

Prevent module resolution conflicts during IndexTTS-2 initialization

Fixed

Fix IndexTTS-2 import errors caused by conflicting system packages
Fix IndexTTS-2 failing to load when system indextts package is installed
Ensure bundled IndexTTS-2 engine uses correct internal modules
Improve compatibility with different Python environments

[4.9.11] - 2025-09-17

Added

Handle faiss circular import issues gracefully during startup
Prevent AttributeError about Float32 in containerized environments

Fixed

Fix startup crash in Docker environments due to faiss circular import
Fix dependency checker crashing with faiss in Docker containers
Improve compatibility with Docker and OpenSUSE setups

[4.9.10] - 2025-09-17

Added

Consistent with other TTS engines temp file handling

Changed

Better file organization and cleanup

Fixed

Fix IndexTTS-2 using Windows temp folder instead of ComfyUI temp
Fix IndexTTS-2 temp files now created in ComfyUI temp folder

[4.9.9] - 2025-09-17

Changed

Better handling of package version conflicts during setup

Fixed

Fix installation issues with numpy downgrade and requirements parsing
Fix opencv-python causing numpy version downgrades during installation
Fix requirements.txt parsing errors with inline comments
Improve dependency installation reliability and compatibility

[4.9.8] - 2025-09-17

Changed

Preserve RVC GPU acceleration performance
Better numpy version management during installation

Fixed

Fix install script downgrading numpy on Linux systems
Fix faiss-gpu installation downgrading numpy from 2.x to 1.26.4 on Linux
Install faiss-gpu with --no-deps to prevent dependency conflicts

[4.9.7] - 2025-09-17

Added

Truncate emotion audio to 20 seconds to prevent memory crashes

Changed

Improve stability with large emotion reference files
Better memory management for Wav2Vec2-BERT processing

Fixed

Fix IndexTTS-2 crashes with long emotion audio files
Fix IndexTTS-2 OOM errors when using long emotion reference audio

[4.9.6] - 2025-09-16

Fixed

Fix ChatterBox Official 23-Lang language resolution and narrator input issues
Fix ChatterBox Official 23-Lang crashes due to language alias resolution errors
Fix ChatterBox Official 23-Lang ignoring direct audio input from opt_narrator
Fix similar language resolution errors in phonemizer and character parser
Improve compatibility with direct audio input across all TTS engines

[4.9.5] - 2025-09-16

Fixed

Fix missing julius dependency causing Higgs Audio failures
Add julius dependency to install script for proper audio processing support
Fix 'No module named julius' error when using Higgs Audio 2 engine
Resolve dependency issue with RVC demucs components requiring julius
Users should run install.py manually in their ComfyUI environment to get the fix
Addresses issue #82 julius dependency error

[4.9.4] - 2025-09-16

Changed

Improve auto device detection to maintain CUDA performance
Better device handling for subsequent generations after model unload

Fixed

Fix VibeVoice stability and performance issues
Fix crashes during garbage collection after VibeVoice generation
Resolve UnboundLocalError with gc variable in model cleanup
Addresses issue #81 garbage collection and device detection problems

[4.9.3] - 2025-09-16

Added

Add clear explanation of .txt vs .reference.txt priority system

Changed

Improve node tooltips for better user experience
Improve IndexTTS-2 emotion_control tooltip formatting and readability
Better documentation of engine-specific requirements

Fixed

Fix Character Voices node tooltips to clearly explain .txt file requirements
Addresses issue #69 tooltip clarity feedback

[4.9.2] - 2025-09-16

Fixed

Fix VibeVoice output compatibility with RVC voice conversion
Fix VibeVoice audio not working with RVC voice converter
Resolve BFloat16 tensor format issues causing conversion failures
Add defensive handling for future engine compatibility
Resolves issue #47 reported audio processing errors

[4.9.1] - 2025-09-16

Added

Add automatic recovery from CPU offloading corruption
Add new IndexTTS-2 integration workflow example

Changed

Improve model stability for multiple generations

Fixed

Fix IndexTTS-2 device errors and add integration workflow
Fix device mismatch error when reusing IndexTTS-2 after model unload

[4.9.0] - 2025-09-16

Added

Add IndexTTS-2 TTS engine with advanced emotion control
New IndexTTS-2 engine with sophisticated emotion control system
Unified emotion control supporting multiple input methods (audio, text, vectors)
Dynamic text emotion analysis with QwenEmotion AI and contextual templates
Per-character emotion control using [Character:emotion_ref] syntax
8-emotion vector control (Happy, Angry, Sad, Surprised, Afraid, Disgusted, Calm, Melancholic)
Audio reference emotion support including Character Voices integration
Emotion intensity control from neutral to maximum dramatic expression
Complete IndexTTS-2 Emotion Control Guide: docs/IndexTTS2_Emotion_Control_Guide.md

Changed

Advanced caching system for improved performance

[4.8.38] - 2025-09-10

Changed

Cleaner installer output without warnings

Fixed

Fix installer deprecation warning
Better compatibility across Python versions

Removed

Remove pkg_resources deprecation warning during installation

[4.8.37] - 2025-09-10

Changed

Improve Higgs Audio reliability across different environments

Fixed

Fix Higgs Audio compatibility with transformers 4.56.0+
Fix generation errors with newer transformers library versions
Resolve 'missing device argument' TypeError during audio generation
Maintain compatibility with older transformers versions

[4.8.36] - 2025-09-10

Changed

Improve stability for character switching with minimal text

Fixed

Fix ChatterBox Official 23-Lang crashes with very short text segments
Fix generation errors when processing single characters like 'A'
Remove unnecessary crash protection padding for 23-lang model

[4.8.35] - 2025-09-10

Changed

Better handling of existing PyTorch installations

Fixed

Fix critical Windows installation failures
Fix Windows UTF-8 encoding errors preventing installation
Fix PyTorch version parsing crashes on systems with CUDA builds
Fix PyTorch 2.6.0 installation failures with automatic fallback

[4.8.34] - 2025-09-10

Added

Add ComfyUI extra_model_paths.yaml support for shared models
Support shared model directories across multiple ComfyUI installations
Enable shared storage configurations as requested by users

Changed

Improve voice discovery with models/TTS/voices fallback support

Fixed

Fix model downloads always going to default ComfyUI models folder
Reference: GitHub issue #63

[4.8.33] - 2025-09-10

Changed

Improve Higgs Audio stability with latest dependencies

Fixed

Fix Higgs Audio compatibility with transformers 4.46.3+
Fix 'property has no setter' errors when using Higgs Audio
Fix StaticCache compatibility issues with newer transformers
Reference: GitHub issue #64

[4.8.32] - 2025-09-09

Added

Add automatic detection of manual Speaker format in text

Changed

Improve SRT subtitle processing with consistent character mapping
Clean up console output for better user experience
Better fallback handling for invalid voice connections

Fixed

Fix VibeVoice voice handling and improve multi-speaker reliability
Fix voice references not loading correctly in Native Multi-Speaker mode
Add helpful error messages when wrong connection types are used

[4.8.31] - 2025-09-09

Added

Faster generation with optimized parameter combination

Changed

Optimize VibeVoice for better quality and consistency
Improve default settings based on community testing (CFG 3.0, 3 inference steps)
Better quality with less background noise and music hallucinations

Fixed

Fix voice generation producing inconsistent results between parameter configurations

[4.8.30] - 2025-09-09

Fixed

Fix VibeVoice cache and compatibility issues
Fix cache not updating when changing inference steps or CFG values
Fix DynamicCache errors with newer transformers library versions
Improve VibeVoice compatibility with latest transformers updates

[4.8.29] - 2025-09-09

Added

Allow ultra-low inference steps (1-3) for fast generation experiments
Add support for CFG 3.0 + 3 steps combination as reported by users

Changed

Improve VibeVoice generation flexibility and default quality
Change default to Native Multi-Speaker mode for better text processing
Better parameter flexibility for quality/speed experimentation

Removed

Remove CFG scale limit to allow testing high values like 3.0

[4.8.28] - 2025-09-09

Added

Add support for 4-bit model quantization to reduce VRAM usage

Changed

Improve voice reference troubleshooting information

Fixed

Fix VibeVoice character voice loading and quantization issues
Fix character voices not being found in VibeVoice engine
Fix misleading completion messages when using character switching
Better compatibility with audio-only engines like VibeVoice

[4.8.27] - 2025-09-09

Changed

Performance improvements by only applying workarounds when needed

Fixed

Improve numba compatibility handling for Python 3.13
Fix contradictory startup messages about JIT status
Better handling of librosa/numba compatibility issues
Cleaner startup process with single compatibility test
Maintain full compatibility with all Python 3.13 environments

[4.8.26] - 2025-09-08

Fixed

Fix Python 3.13 compatibility issues in conda environments
Fix audio hashing errors with Python 3.13 + conda/miniconda
Resolve 'get_call_template' compilation errors
Improve numba compatibility across different Python environments

[4.8.25] - 2025-09-07

Added

Note: System RAM still accumulates when switching modes

Fixed

Fix VibeVoice VRAM clearing
Fix VRAM clearing to prevent out of memory errors
Fix models not switching when changing attention modes

[4.8.24] - 2025-09-07

Changed

Improve model caching to properly handle configuration changes

Fixed

Fix VibeVoice model reloading and improve stability
Fix VibeVoice not reloading when changing attention mode settings
Resolve transformers compatibility issues with VibeVoice engine
Fix numba JIT disable logic showing wrong Python version messages

Removed

Remove excessive debug output for cleaner console logs

[4.8.23] - 2025-09-07

Added

Add ComfyUI VRAM management support for ChatterBox Official 23-Lang
Add ComfyUI VRAM management support (Clear VRAM button now works)

Changed

Improve memory efficiency for ChatterBox Official 23-Lang engine

Fixed

Fix model caching to prevent reloading on each generation
Fix Python 3.13 compatibility issues with audio processing

[4.8.22] - 2025-09-07

Added

Add ComfyUI VRAM management support (Clear VRAM button now works)

Changed

Improve model caching to reuse loaded models across generations
Better memory efficiency for VibeVoice engine

Fixed

Fix VibeVoice memory issues and add VRAM management
Fix model reloading on each generation causing out-of-memory errors

[4.8.21] - 2025-09-06

Added

Add proper null checks for ChatterBox factory functions

Fixed

Fix remaining voice conversion import issues on Python 3.13
Fix voice conversion still failing after import success
Improve error handling in unified model interface
Resolve edge case causing 'NoneType' errors despite successful imports

[4.8.20] - 2025-09-06

Changed

Change default vocal separation model to faster, higher-quality option

Fixed

Fix Python 3.13 compatibility for RVC and vocal separation
Fix RVC voice conversion working fully in Python 3.13 environments
Fix vocal separation compatibility with most UVR models in Python 3.13
Fix missing RVC f0 extraction methods (harvest, dio, fcpe) now available
Fix model download and path resolution issues for specialty models
Improve error messages with clear guidance instead of technical dumps

[4.8.19] - 2025-09-06

Added

Restore complete RVC functionality that was working in previous Python versions

Fixed

Fix RVC voice conversion compatibility in Python 3.13
Fix RVC voice conversion failing in Python 3.13 environments
Fix missing f0 extraction methods (pm, harvest, dio, fcpe) that were broken
Fix RVC model loading and pitch extraction reliability issues
Fix Python 3.13 numba/librosa dependency conflicts causing crashes

[4.8.18] - 2025-09-06

Changed

Improve ChatterBox support for incomplete language models
Better fallback handling for ChatterBox multilingual TTS generation

Fixed

Fix ChatterBox French TTS model loading issues
Fix ChatterBox French model failing to load due to missing tokenizer

[4.8.17] - 2025-09-06

Fixed

Fix voice conversion crashes on Python 3.13 systems
Fix voice changer failing with 'NoneType' object error
Improve Python 3.13 compatibility for voice conversion
Better error handling when watermarking library is unavailable
Resolve custom model loading issues (e.g. local:German-KB-H2-Merge-v2)

[4.8.16] - 2025-09-06

Added

Add manual implementations as fallbacks for removed librosa utilities
RVC should now work with any librosa version including latest releases

Fixed

Improve RVC compatibility with all librosa library versions
Fix additional librosa compatibility issues with newest versions
Handle missing pad_center, tiny, and normalize functions individually

[4.8.15] - 2025-09-06

Added

Support both old and new librosa library versions automatically
RVC voice conversion now works regardless of librosa version

Fixed

Fix RVC voice conversion failing with newer librosa versions
Fix RVC failing with 'cannot import name pad_center from librosa.util' error
Resolve issues caused by manual dependency updates bypassing our install script

[4.8.14] - 2025-09-06

Added

Both TTS engines now work reliably with latest transformers versions

Fixed

Fix compatibility with latest transformers library versions
Fix VibeVoice failing with 'DynamicCache' object has no attribute 'key_cache' error
Fix Higgs Audio 2 failing with 'StaticCache' object has no attribute 'key_cache' error
Support transformers 4.56+ while maintain backward compatibility
Resolve generation failures after clean ComfyUI installations with newer dependencies

[4.8.13] - 2025-09-06

Added

Ensure all audio processing components work correctly with latest Python

Fixed

Fix Python 3.13 compatibility issues with ChatterBox Official 23-Lang engine
Fix crashes when using ChatterBox Official 23-Lang with Python 3.13
Resolve numba compilation errors that prevented voice generation
Improve compatibility with modern Python environments

[4.8.12] - 2025-09-06

Changed

Improve parameter flow consistency across all F5-TTS components

Fixed

Fix F5-TTS phonemization toggle not working properly
Fix toggle not disabling phonemization for non-English models
Fix cache not invalidating when changing phonemization setting

Removed

Remove debug messages when phonemization is disabled

[4.8.11] - 2025-09-06

Added

Add support for RTX 50 series GPUs (Blackwell architecture)
Enhanced input validation prevents generation failures
Improved audio preprocessing reduces weird sound artifacts

Changed

Better handling of corrupted audio inputs that could cause noise generation

Fixed

Improve VibeVoice reliability and GPU compatibility

[4.8.10] - 2025-09-06

Added

ChatterBox Official 23-Lang now recognizes existing safetensors files
Prevents unnecessary re-downloads when safetensors already present

Fixed

Fix download detection for existing safetensors files

[4.8.9] - 2025-09-06

Added

Support both .pt and safetensors formats automatically
Add fallback download if official format unavailable

Fixed

Fix ChatterBox Official 23-Lang safetensors compatibility
New users get official .pt files for guaranteed compatibility
Users with safetensors get compatibility warnings but keep working
Fix missing model files triggering proper re-downloads

[4.8.8] - 2025-09-06

Added

Add phonemization control toggle in F5-TTS Engine for user customization
Disable phonemization for Portuguese F5-PT-BR (was making quality worse)
Add experimental phonemization with user feedback prompts for other languages

Changed

Better debugging messages when phonemization is actually used

Fixed

Fix F5-TTS multilingual support with phonemization control
Fix Polish and other non-English F5-TTS model text processing
Fix phonemization working consistently across both TTS and SRT workflows

[4.8.7] - 2025-09-06

Fixed

Fix ChatterBox crashes with newer transformer versions
Fix crashes when using ChatterBox with transformers 4.36+
Fix ChatterBox Official 23-Lang model failing to generate
Resolve attention implementation compatibility issues

[4.8.6] - 2025-09-05

Added

VibeVoice now properly respects chunk_minutes setting
chunk_minutes=0 now correctly disables all chunking (fixes #38)

Fixed

Fix VibeVoice chunking controls not working correctly
Fix disabling chunking having no effect until restart
Fix chunk size changes requiring ComfyUI restart

[4.8.5] - 2025-09-05

Added

Support model names containing special characters like colons
Voice conversion now works properly with all custom model names

Fixed

Fix voice conversion failing with custom local models
Fix voice conversion crashes when using local models with custom names
Better error handling for local model detection

[4.8.4] - 2025-09-05

Added

Parameters now properly affect audio generation instead of using stale cached results
SRT generation now responds correctly to all parameter changes in real-time

Fixed

Fix cache not updating when changing ChatterBox Official 23-Lang parameters
Fix cache invalidation not working when changing TTS parameters like temperature or exaggeration
Resolve issue where changing advanced parameters (repetition_penalty, min_p, top_p) had no effect

[4.8.3] - 2025-09-05

Changed

Improve audio processing reliability for complex SRT files
Better handling of multilingual character switching with pause tags

Fixed

Fix SRT overlapping and character voice issues
Fix SRT overlapping mode not working properly (audio was being cut off)
Fix cache not updating when changing ChatterBox Official 23-Lang parameters
Fix character voices reverting to narrator after pause tags like [pause:10s]

[4.8.2] - 2025-09-05

Added

All timing modes now work properly with character switching and overlaps

Changed

Better audio processing for pause tags and mixed audio formats

Fixed

Fix SRT overlapping and caching issues
Fix overlapping subtitles not playing simultaneously across all TTS engines
Fix audio parameter changes not invalidating cache in ChatterBox Official 23-Lang
Fix crashes when using concatenate and stretch-to-fit timing modes

[4.8.1] - 2025-09-05

Added

All TTS engines now handle overlapping dialogue correctly

Changed

Improve timing accuracy for conversations with multiple speakers
Better audio synchronization for complex subtitle files

Fixed

Fix SRT audio overlapping not working properly
Fix overlapping subtitles being cut off instead of playing simultaneously

[4.8.0] - 2025-09-05

Added

Add ChatterBox Official 23-Lang multilingual TTS engine
New advanced text-to-speech engine with support for 23 languages
Generate high-quality speech in Arabic, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Italian, Japanese, Korean, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swedish, Swahili, and Turkish
SRT subtitle timing support for synchronized audio generation
Voice conversion capabilities supporting all 23 languages
Improved audio quality and consistency across all features

Changed

Better voice reference processing and character switching

Fixed

Fix audio pitch issues that made voices sound unnaturally deep

[4.7.2] - 2025-09-04

Added

Voice conversion now works properly with custom local models
Fixes #36

Fixed

Fix voice conversion failing on Windows with local models
Fix Windows path errors when using local ChatterBox models
Better cross-platform compatibility for model loading

[4.7.1] - 2025-09-04

Added

Add backup model sources for reliable downloads
Add SageAttention GPU-optimized attention for 2-4x speed boost
SageAttention implementation based on https://github.com/wildminder/ComfyUI-VibeVoice

Changed

Better GPU memory optimization with mixed-precision kernels

Fixed

Fix VibeVoice compatibility and performance improvements
Fix compatibility with newer transformers library versions
Fix tokenizer downloads to avoid cache conflicts

[4.7.0] - 2025-09-04

Added

Add ChatterBox support for 9 new languages
NEW: Italian bilingual model with automatic [it] prefix for Italian text
NEW: French model with 1400+ hours training and voice cloning
NEW: Russian complete model with training artifacts
NEW: Armenian and Georgian models with unique architectures
NEW: Japanese and Korean support with shared components
NEW: Three German variants (Standard, Best quality, Expressive with emotion tags)

[4.6.28] - 2025-09-03

Added

Add universal audio format support for VideoHelper LazyAudioMap objects
All audio-processing nodes now work with VideoHelper audio without workarounds

Fixed

Fix VideoHelper Suite audio compatibility
Fix Voice Changer not detecting audio from VideoHelper Suite Load Video nodes
Improve compatibility with Character Voices and standard AUDIO inputs

[4.6.27] - 2025-09-03

Changed

Improve installation reliability across different Python versions
Better handling of audio separation dependencies

Fixed

Fix numpy compatibility issues with other ComfyUI nodes
Support numpy 1.26.4 for better compatibility with other custom nodes
Fix VibeVoice installation failing due to missing Microsoft repository (for whataver reason the repo is now gone)

[4.6.26] - 2025-09-01

Added

Hardware settings grouped at top (device, quantization, attention mode)
Quality controls placed together for intuitive workflow
More logical parameter flow from hardware to generation settings

Changed

Improve VibeVoice Engine parameter layout
Better parameter organization for easier configuration

[4.6.25] - 2025-09-01

Added

Add 4-bit quantization option for VRAM savings (helpful when VRAM is limited)
Add inference steps control for quality vs speed adjustment

Changed

Enhance VibeVoice engine with advanced parameters and fixes
Add attention mode selection for performance optimization
Realistic performance expectations in parameter descriptions

Fixed

Fix compatibility issues with newer Transformers library versions
Better device handling and error messages for quantization

[4.6.24] - 2025-09-01

Added

Maintain same functionality with improved model source
Backward compatible with existing model installations

Changed

Update VibeVoice 7B to official Microsoft repository
Use official Microsoft repository for better reliability

[4.6.23] - 2025-09-01

Fixed

Fix VibeVoice engine showing as available when package not installed
Fix misleading init messages for VibeVoice engine
Improve dependency checking consistency across all engines
Prevent engine from appearing in node list if dependencies missing

[4.6.22] - 2025-09-01

Changed

Improve clarity when using custom downloaded language models
Better support for unidentified local language models

Fixed

Fix local F5-TTS model language display issue
Fix confusing language code display for local F5-TTS models

[4.6.21] - 2025-09-01

Added

Add clear warnings that RVC is for voice conversion only (not text-to-speech)

Changed

Improve RVC voice conversion clarity and connections

Fixed

Fix RVC Engine connection suggestions in ComfyUI interface
Improve node connection compatibility between RVC and Voice Changer

[4.6.20] - 2025-09-01

Changed

Improve download reliability with fallback sources

Fixed

Fix RVC content-vec-best model failing to auto-download
Fix voice conversion model download issues
Add automatic format conversion for compatibility

[4.6.19] - 2025-08-31

Added

Untagged text now correctly uses the engine's selected model

Changed

Only tagged characters or alias-mapped characters change models

Fixed

Fix F5-TTS not using selected model and language issues
Fix F5-TTS generation failing due to variable scope error
Fix F5-TTS SRT ignoring user's model selection (e.g., Hindi model)

[4.6.18] - 2025-08-31

Added

Add automatic fallback to alternative audio processing when needed

Fixed

Fix VibeVoice failing with numba compatibility error
Fix audio loading error preventing VibeVoice from working on some systems
Improve compatibility with different Python environments

[4.6.17] - 2025-08-30

Changed

Improve text processing for longer audio generation

Fixed

Fix VibeVoice audio generation failing with longer text
Fix chunking error preventing VibeVoice from generating audio over 30 seconds
Resolve issue where chunking would fail even when disabled

[4.6.16] - 2025-08-30

Added

More reliable model unloading when using 'Clear VRAM' button
Enhanced stability when switching between different models

Changed

Improve VibeVoice memory management and unloading
Better integration with ComfyUI's memory management system
Consistent architecture across all TTS engines for better maintainability

[4.6.15] - 2025-08-30

Added

VibeVoice models now properly unload when clicking 'Unload Models' button
VRAM is actually freed when unloading (not just cache cleared)
Models are properly recreated fresh after unload

Fixed

Fix VibeVoice unload button not freeing VRAM
Fix crash when generating after unloading models

[4.6.14] - 2025-08-30

Added

Add support for existing VibeVoice installations and HuggingFace cache
Eliminates need to restart ComfyUI between VibeVoice generations

Changed

VibeVoice models now stay loaded between uses for better performance

Fixed

Fix VibeVoice memory errors and model loading failures
Fix allocation errors when generating audio multiple times
Fix VibeVoice-7B model failing to download (404 errors)

[4.6.13] - 2025-08-30

Added

Add support for Kijai's MelBandRoFormer models (.safetensors)

Changed

Improve model-specific output handling and inversion detection

Fixed

Fix MelBandRoFormer models now working properly
Fix denoise_mel_band_roformer models (27.99 SDR) compatibility
Fix architecture mismatch errors preventing model loading

[4.6.12] - 2025-08-30

Added

Add VibeVoice legacy path and cache support
Support models from original VibeVoice-ComfyUI extension
Reuse HuggingFace cached VibeVoice models when available
Avoid re-downloading models for users upgrading from other extensions

Changed

Improve model loading efficiency with multi-source fallback

[4.6.11] - 2025-08-30

Added

Long text generation now works properly with Higgs Audio engine

Fixed

Fix Higgs Audio TTS chunking error with long text
Fix crash when using Higgs Audio with text that requires chunking
Resolve sample_rate parameter error in audio combination

[4.6.10] - 2025-08-30

Changed

Update project description to mention all supported TTS engines.

Removed

Remove virtual environment files and update project info
Remove accidentally committed virtual environment files.

[4.6.9] - 2025-08-30

Added

Add comprehensive VibeVoice documentation to README.

Changed

Clean up debug output and improve documentation

Removed

Remove verbose debug logs for cleaner console output.

[4.6.8] - 2025-08-30

Added

Implement comprehensive solution for module import issues
Comprehensive solution for 'No module named' errors across all setups

Changed

Smart project root detection and multi-strategy import fallbacks
Zero performance impact on working installations

Fixed

Add robust import system for cross-environment compatibility
Fix module discovery issues in conda, Windows, and other environments

[4.6.7] - 2025-08-30

Added

Add manual Speaker format support and voice priority system
Support both [Character] tags and manual 'Speaker 1: Hello' format in Native Multi-Speaker mode
Add voice priority system: connected speaker2/3/4 inputs override character aliases with warnings
Support up to 4 speakers with flexible voice assignment options

Changed

Improve user experience with clear format explanations and debug logging

Fixed

Fix Speaker 1 voice mapping to use main narrator correctly

[4.6.6] - 2025-08-29

Added

Complete VibeVoice SRT subtitle support
Add Custom Character Switching mode for detailed character control
Support all timing modes (stretch, pad, natural, concatenate)

Changed

Add Native Multi-Speaker mode for efficient multi-character subtitles
Improve cache management when switching between modes

Fixed

Fix character voice cloning in SRT processing

[4.6.5] - 2025-08-29

Added

Complete VibeVoice SRT implementation with character voices
VibeVoice SRT now uses proper character voice files
Alice and Bob characters now sound like their voice samples

Changed

Improved voice reference handling for better audio quality

Fixed

Fix character voices not working in SRT mode

[4.6.4] - 2025-08-29

Added

Ensure cache invalidation triggers properly in all Python environments

Fixed

Fix module import errors in conda environments
Fix ModuleNotFoundError not being caught in import fallback logic
Improve conda environment compatibility for ChatterBox nodes
Complete fix for 'No module named utils.audio.audio_hash' error

[4.6.3] - 2025-08-29

Changed

Improve module import reliability for conda and isolated environments
Improve robustness across different Python environment setups
Zero performance impact on working installations

Fixed

Enhance fix for 'No module named utils.audio.audio_hash' error
Add conda environment compatibility for import issues
Better error handling for lazy imports in ChatterBox nodes

[4.6.2] - 2025-08-29

Added

ChatterBox TTS and SRT nodes now work reliably for all installation types

Changed

Improve robustness of lazy imports across various Python environments

Fixed

Fix module import errors in ChatterBox TTS and SRT nodes
Fix 'No module named utils.audio.audio_hash' error in TTS generation
Resolve module discovery issues for users with different ComfyUI setups

[4.6.1] - 2025-08-29

Added

Add support for .safetensors format in Voice Conversion
German and Norwegian voice conversion now works properly

Fixed

Fix Voice Changer support for German and Norwegian models
Fix Voice Changer failing to load German/Norwegian models
Improve compatibility with multilingual ChatterBox models

[4.6.0] - 2025-08-29

Added

Add VibeVoice TTS engine with long-form generation
New VibeVoice TTS engine provides advanced multi-speaker synthesis:
Support for Microsoft VibeVoice 1.5B and 7B models
Generate up to 90 minutes of continuous audio
Native multi-speaker mode supports up to 4 distinct voices
Custom character switching with [CharacterName] tag support
Time-based chunking interface for managing long-form content
Pause tag support with [pause:Xs] syntax
Seamless integration with existing Character Voices workflow
Technical features:
Unified cache system with proper character isolation
High-quality audio processing matching official implementation
Full integration with existing TTS Text and Character Voices nodes
Note: This is an initial implementation focused on core functionality.
SRT subtitle support and additional features will be added in upcoming releases.

Fixed

Automatic model downloading and dependency management

[4.5.30] - 2025-08-28

Added

Support case variations in model names

Fixed

Fix F5-TTS model compatibility issues
Fix F5TTS_V1_Base model name validation error
Improve backward compatibility with existing workflows

[4.5.29] - 2025-08-28

Added

Clearer messaging when local vs cloud models are used

Changed

Improve F5-TTS local model support and reduce Google dependencies
Better local model detection for E2-TTS variants

Fixed

Fix F5-TTS to use local models when available without requiring google-cloud-storage
Reduce dependency requirements for users with manual model downloads

[4.5.28] - 2025-08-28

Changed

Better audio package installation for Mac users

Fixed

Fix Mac M1 compatibility and installation issues
Fix architecture mismatch errors on M1/M2 Macs
Resolve wandb circular import affecting all TTS nodes
Add automatic system dependency checking for macOS
Improve ARM64 package compatibility on Apple Silicon

[4.5.27] - 2025-08-28

Added

Enable safe VRAM management in Memory Safe mode

Changed

Add CUDA Graph toggle: High Performance (55+ tok/s) vs Memory Safe (12 tok/s)

Fixed

Fix Higgs Audio model unloading crashes with performance toggle
Fix crashes when using 'Unload Models' with Higgs Audio engine
Fix multiple generation cycles and cache state issues
Resolve fundamental CUDA Graph vs ComfyUI memory management conflict

[4.5.26] - 2025-08-27

Fixed

Fix Higgs Audio advanced RAS parameter support
Fix parameter cache invalidation for proper audio regeneration
Fix configuration flow ensuring parameters reach underlying model correctly

Added

Add missing RAS (Repetition Avoidance Sampling) controls to Higgs Audio engine

Changed

Improve speech quality control with force_audio_gen and repetition settings

[4.5.25] - 2025-08-27

Fixed

Fix Silent Speech Analyzer preview not reflecting post-processing parameters
Fix cache invalidation causing unnecessary video re-analysis when adjusting merge threshold

Changed

Improve Silent Speech Analyzer performance with optimized caching system

[4.5.24] - 2025-08-27

Added

Fix F5-TTS Speech Editor workflow compatibility with v4 unified architecture

[4.5.23] - 2025-08-27

Added

Fix F5-TTS speech editing issues with E2TTS models and vocabulary handling

Resolves compatibility problems when using E2TTS models with the F5-TTS speech editor.

Speech editing now works correctly with all supported F5-TTS model variants including:
Fixed vocab size mismatch issues between different model types
Integrated edit engine with unified model interface
Eliminated duplicate model loading in edit operations
Improved error handling for missing vocabulary files

[4.5.22] - 2025-08-27

Added

Add automatic detection for missing Linux system dependencies with helpful error messages and install instructions

[4.5.21] - 2025-08-27

Added

Add new Unified 📺 TTS SRT workflow showcasing all engines and features, remove outdated individual engine workflows

[4.5.20] - 2025-08-26

Added

Enhanced Higgs Audio 2 language processing for better multilingual support

Unlike ChatterBox/F5-TTS which use separate models per language, Higgs Audio 2's base model supports multiple languages natively. Improved language tag processing:

Smart language hint conversion: [En:Alice] → [English] Hello there for better model context
Explicit language detection: Only adds hints when user specifies language prefix (not character defaults)
Character switching preservation: Maintains proper voice assignment while converting language tags
Console logging improvements: Shows actual processed text sent to engine instead of raw SRT content
Support for all language variations: English, German, Norwegian, French, Spanish, Portuguese, etc.

Result: Higgs Audio 2 receives meaningful language context for optimal multilingual performance while keeping processing clean and efficient.

[4.5.19] - 2025-08-26

Added

Fix numba version compatibility by upgrading to 0.61.2+ for NumPy 2.2+ support, resolving engine loading failures

[4.5.18] - 2025-08-26

Changed

Fix Higgs Audio 2 Engine Character Voices compatibility

• Higgs Audio 2 Engine now accepts 🎭 Character Voices as secondary narrator input • Native multi-speaker modes properly use reference text from both Character Voices
• Improved voice cloning quality for SPEAKER1 in native multi-speaker conversations • Secondary narrator now uses dedicated reference text instead of sharing primary narrator text

[4.5.17] - 2025-08-26

Fixed

Fix F5-TTS narrator segments using wrong language when selecting non-English models

[4.5.16] - 2025-08-26

Changed

Remove excessive debug messages and consolidate redundant logs for cleaner user experience

Cleaned up verbose logging in:

ChatterBox engine: Consolidated model loading messages, removed cache debug spam
F5-TTS engine: Silenced path verbosity, fixed Vocos local detection, removed internal logs
Higgs Audio engine: Commented out boson_multimodal INFO logs, unified completion messages
Universal nodes: Removed engine creation debug messages
Smart Loader: Silenced routine cache operations
Character Voices: Consolidated multiple setup messages into single line
Model management: Removed technical integration details

Result: Significantly reduced console noise while preserving important error and completion messages

[4.5.15] - 2025-08-26

Changed

Fix PyTorch installation failing with CUDA systems and improve compatibility detection for different GPU configurations

[4.5.14] - 2025-08-26

Added

Fix critical 'NoneType' object is not callable error by implementing lazy watermarker initialization.
Watermarking now disabled by default with graceful fallback if initialization fails, eliminating model loading failures in Python 3.13 environments

[4.5.13] - 2025-08-26

Changed

Remove unnecessary debug messages from startup logs - cleaned up model factory registration spam, SRT module loading messages, and redundant status messages to provide cleaner startup experience while keeping essential error reporting and progress indicators

[4.5.12] - 2025-08-26

Added

Fix 'NoneType' object is not callable error when loading ChatterBox TTS models in Python 3.13 Windows environments by adding Unicode encoding error handling to import_manager

[4.5.11] - 2025-08-25

Added

Implement ComfyUI memory management integration for TTS models

Add ComfyUI model management support:

ChatterBox and F5-TTS models can now be unloaded via ComfyUI Manager
Fix device assignment after memory cycles
Clean up debug message spam
Reduce repetitive Higgs Audio warnings

Note: Higgs Audio models cannot be unloaded due to CUDA graph limitations

Relates to #6

[4.5.10] - 2025-08-25

Added

Fix Higgs Audio seed parameter not affecting cache key - changing seeds now properly regenerates audio instead of hitting cache

Implement comprehensive ComfyUI model management integration:

All TTS models now integrate with ComfyUI's native model management system
'Clear VRAM' and 'Unload models' buttons now work with TTS models for automatic memory management
Create unified model loading interface standardizing all engines (ChatterBox, F5-TTS, Higgs Audio, RVC, Audio Separation)
Replace static model caches with ComfyUI-managed dynamic loading
Add ComfyUI-compatible model wrapper enabling proper integration with ComfyUI's memory management
Standardize factory pattern for model creation across all engines
Enhance model fallback utilities with generic local-first loading behavior

[4.5.9] - 2025-08-25

Fixed

Fix memory allocation issues when running Higgs Audio after other TTS models by reducing KV cache sizes from [2048, 8192, 16384] to [1024, 2048, 4096].
This prevents out-of-memory errors on 24GB GPUs while maintaining full functionality for typical TTS usage.

[4.5.8] - 2025-08-25

Added

Complete Python 3.13 compatibility by implementing comprehensive librosa fallback system for ChatterBox engine.
Replaces all librosa calls with safe fallbacks that try librosa first (for quality) then torchaudio (for Python 3.13 compatibility).
Fixes numba compilation errors in voice encoder melspec module.

[4.5.7] - 2025-08-25

Fixed

Python 3.13 compatibility: Higgs Audio now works on Python 3.13 by using torchaudio instead of librosa (no quality impact, better performance)

[4.5.6] - 2025-08-24

Fixed

Complete comprehensive HuggingFace cache detection system across all model downloaders - prevents duplicate downloads by checking local files, HuggingFace cache, then downloading to local (never to cache).
Fix HuBERT model download structure that was preventing downloads from completing successfully.

[4.5.5] - 2025-08-22

Fixed

Fix critical protobuf dependency conflict preventing ChatterBox, F5-TTS, and Higgs Audio engines from loading.
Move descript-audiotools to --no-deps installation to preserve protobuf 6.x compatibility.
All 19 nodes now load successfully on Python 3.13.6.

[4.5.4] - 2025-08-22

Added

Fix Windows Unicode encoding error preventing installation on clean systems Fix UnicodeEncodeError when Windows console cannot display emoji characters Add graceful fallback to text-based logging for Windows CP1252 encoding Resolve install script crash that prevented ComfyUI Manager installations

[4.5.3] - 2025-08-22

Added

Add comprehensive installation system with intelligent dependency management Add Python 3.13 full compatibility support with MediaPipe to OpenSeeFace fallback Add intelligent install.py script with automatic conflict resolution Add environment detection and safety warnings for system Python usage Add NumPy version constraints to prevent Numba compatibility issues Add automatic RVC dependencies installation support Update requirements.txt with comprehensive dependency documentation Update README with detailed installation instructions and Python 3.13 compatibility notes Add dedicated Higgs Audio 2 model installation section with complete folder structure Verify ComfyUI Manager integration with automatic install.py execution Fix all bundled engines compatibility: ChatterBox, F5-TTS, Higgs Audio, RVC

[4.5.2] - 2025-08-22

Changed

Fix Higgs Audio engine compatibility with transformers 4.46+ including attention API changes, cache handling, and generation loop updates

[4.5.1] - 2025-08-21

Added

Fix critical Higgs Audio download and loading issues
Fix sharded model file downloads (model-00001-of-00003.safetensors, etc.)
Add missing tokenizer files (tokenizer.json, tokenizer_config.json, special_tokens_map.json) to model directory
Fix tokenizer loading from organized local structure instead of HuggingFace cache
Add proper completeness validation for downloaded model directories
Implement fallback tokenizer loading with LlamaTokenizer for custom configs
Ensure all Higgs Audio models load from organized TTS/HiggsAudio/ structure without cache duplication

[4.5.0] - 2025-08-21

Added

🎙️ Higgs Audio 2 Voice Cloning Integration

🌟 Major New Features

🎙️ Higgs Audio 2 Voice Cloning Engine

State-of-the-art voice cloning from 30+ second reference audio samples
Multi-speaker conversation support with seamless character switching
Real-time voice replication with exceptional audio quality
Universal integration - works with existing TTS Text and TTS SRT nodes
Advanced generation controls - fine-tune temperature, top-p, top-k, and token limits
Multi-language capabilities - English (tested), with potential support for Chinese, Korean, German, and Spanish

🧠 Smart Audio Processing System

Intelligent chunk combination with automatic boundary detection
Per-junction analysis for optimal audio segment merging
Universal implementation across all TTS engines (ChatterBox, F5-TTS, Higgs Audio)
Enhanced audio quality through smart text structure analysis

🏗️ Architecture Improvements

⚙️ Modular Processor Architecture

Complete architectural consistency across all TTS engines
Unified delegation pattern - clean separation between user interface and engine processing
HiggsAudio SRT and TTS processors following established patterns
Simplified maintenance with modular, testable components

🔄 Enhanced Processing Pipeline

Modular SRT overlap detection across all engines
Centralized cache management with improved invalidation
Smart model loading and initialization optimizations
Progress tracking with real-time feedback for long operations

🎯 User Experience Enhancements

📊 Improved Feedback and Documentation

Real-time progress bars for Higgs Audio generation
Enhanced tooltips with detailed parameter explanations
Optimized parameter ranges for better generation control
Voice reference management with flexible discovery system

🚀 Performance Optimizations

Instant cache regeneration for previously processed content
Automatic model management with organized TTS/HiggsAudio/ structure
Memory-efficient processing with smart resource utilization
Seamless audio combination eliminating artifacts between segments

🔧 Technical Improvements

🎵 Audio Quality Enhancements

Advanced chunking algorithms with sentence boundary detection
Intelligent silence insertion between audio segments
Cache-optimized processing maintaining quality while improving speed
Professional-grade audio output across all supported engines

🌐 Integration and Compatibility

Voice cloning compatibility with existing character switching system
Reference audio support for custom voice creation
Multi-engine harmony - all engines work consistently with unified nodes
Backward compatibility maintained for existing workflows

[4.4.0] - 2025-08-16

Added

Add new 🗣️ Silent Speech Analyzer node for video mouth movement analysis Features experimental viseme detection for vowels (A, E, I, O, U) and consonants (B, F, M, etc.) Provides 3-level analysis: frame detection → syllable grouping → word prediction Generates base SRT timing files for manual editing and use with TTS SRT nodes Includes MediaPipe integration for production-ready mouth movement tracking Supports visual feedback in preview videos with detection overlays Word predictions use CMU Pronouncing Dictionary (135K+ words) as phonetic placeholders Optimized default values for better detection sensitivity and response time Note: OpenSeeFace provider available but experimental - MediaPipe recommended Important: Results are experimental approximations requiring manual editing

[4.3.7] - 2025-08-13

Added

Implement TTS/ structure with legacy support for cleaner model folder organization
Add native safetensors support for Japanese and Korean Hubert models
Update all engines (Chatterbox, F5-TTS, RVC, UVR) to use organized TTS/ paths

Fixed

Fix RVC Hubert model compatibility issues with automatic .pt to .safetensors conversion
Fix misleading hubert-base-rvc model that failed but claimed to be recommended
Ensure models download to clean TTS/ folder structure instead of cluttered root

[4.3.6] - 2025-08-13

Fixed

Remove duplicate models (hubert-base, hubert-soft) and invalid wav2vec2 model
Fix failing download URLs for HuBERT models
Consolidate to 5 authentic HuBERT variants from verified sources

[4.3.5] - 2025-08-13

Added

Add HuBERT model selection to RVC Engine with 8 different model options
Add auto-download support and language-specific recommendations for HuBERT models
Add comprehensive tooltips explaining each HuBERT model

Fixed

Fix broken URLs for HuBERT model downloads

[4.3.4] - 2025-08-13

Added

Add comprehensive workflow documentation including new unified RVC+ChatterBox workflow

Changed

Improve RVC Engine UI by removing duplicate pitch_detection parameter
Improve RVC Pitch Options with enhanced sliders
Better separation of concerns between engine configuration and detailed pitch extraction options

[4.3.3] - 2025-08-13

Fixed

Fix RVC model dropdown to show both downloadable and local models like F5-TTS
Load RVC Character Model node now properly displays both model types in same dropdown
Add 'local:' prefix for local models to distinguish from downloadable models
Match F5-TTS dropdown behavior for consistent user experience across engines

[4.3.2] - 2025-08-13

Added

Add comprehensive RVC models setup guide to README documentation
Add detailed section explaining RVC model auto-download system and folder structure
Add guidance on available character models and download locations

Fixed

Fix Load RVC Character Model node to show correct downloadable models instead of generic fallbacks

[4.3.1] - 2025-08-13

Fixed

Fix RVC vocal removal node failing to load due to missing ffmpeg-python dependency
Replace ffmpeg-python package usage with direct subprocess calls to system ffmpeg binary
Eliminate dependency requirement by matching existing SRT timing implementation approach

[4.3.0] - 2025-08-12

Added

Version 4.3.0 - Architecture Modernization

🏗️ Core Architecture

Universal Streaming Infrastructure: Complete architectural overhaul creating extensible framework for future engines
Smart Model Loading: Prevents duplicate model instances, eliminating memory exhaustion when switching between processing modes
Thread-Safe Design: Stateless wrapper architecture eliminates shared state corruption risks

⚙️ Engine Extensibility

Universal Adapter System: Standardized framework ready for future TTS engines
Modular Design: Clean separation between engine logic and processing coordination

📝 Performance Notes

Sequential Processing Recommended: Despite parallel processing infrastructure, batch_size=0 (sequential) remains optimal for performance
Memory Efficiency: Improved model sharing between traditional and streaming modes

🔧 Technical Improvements

Centralized Caching: Content-based hashing system for reliable cache consistency
Code Cleanup: Removed unused experimental processor code

⚠️ Known Issues

Console Logging: Needs further cleanup (more verbose than previous versions)
Parallel Processing: Available but slower than sequential - use batch_size=0 for best performance

This version focuses on architectural foundation and extensibility rather than immediate performance gains. The streaming infrastructure provides a robust base for future development while maintaining compatibility with existing workflows.

[4.2.3] - 2025-08-08

Fixed

Fix character alias language resolution bug preventing character voice switching, improve console logging with caching to eliminate spam, fix pause tag parsing regex, and update workflow status in README

[4.2.2] - 2025-08-08

Fixed

Remove IndicF5-Hindi model support due to architecture incompatibility

IndicF5 uses custom transformer layers that are incompatible with F5-TTS DiT architecture. Alternative: F5-Hindi-Small remains available for Hindi TTS. Other Indian languages now fall back to base F5TTS models.

[4.2.2] - 2025-08-08

Removed

REMOVED: IndicF5-Hindi model support due to fundamental architecture incompatibility
- IndicF5 uses custom transformer layers (time_embed.time_mlp, text_embed.text_blocks) that are incompatible with F5-TTS DiT architecture
- Model weights cannot be loaded into standard F5-TTS pipeline due to layer structure differences
- Attempting integration would require extensive custom DiT implementation beyond scope of F5-TTS integration
- Alternative: F5-Hindi-Small (SPRINGLab/F5-Hindi-24KHz) remains available for Hindi TTS (632MB, fully compatible)
- Impact: Indian languages (Assamese, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu) now fall back to base F5TTS models

Technical Notes

Removed IndicF5Engine, all related imports, and model configuration entries
Updated language mapping to use F5-Hindi-Small for Hindi, base models for other Indian languages
Cleaned up documentation and README references

[4.2.1] - 2025-08-08

Added

Add comprehensive Hindi support with F5-Hindi-Small (632MB) and IndicF5-Hindi (1.4GB) models supporting 11 Indian languages
Add support for all 11 Indian languages (Hindi, Assamese, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu) through IndicF5 multilingual model
Add flexible Small vs Base architecture detection that works with any future Small model variants

Fixed

Fix language switching logic to properly detect explicit language tags in original text
Fix model loading bugs with enhanced architecture detection for Small models (18 layers vs Base 22 layers)
Fix language mapping system using existing language_mapper integration instead of hardcoded mappings

[4.2.0] - 2025-08-08

Added

Add advanced RVC voice conversion and vocal separation capabilities

🎵 Enhanced RVC Features:

Advanced RVC parameter controls matching Replay terminology (pitch, pitch_detection, index_ratio, etc.)
Comprehensive audio stem extraction for voice, echo, and noise isolation
New merge audio node for sophisticated audio blending
Clean separation of concerns between RVC Engine and RVC Pitch Options nodes

🔧 Audio Processing Improvements:

Vocal removal with adjustable aggressiveness parameter
Enhanced model architecture detection and compatibility
Robust audio format handling and processing pipeline
Multiple new vocal separation models with performance ratings

⚠️ Experimental Features:

SCNet SOTA architecture implementation (10.08 SDR) - EXPERIMENTAL with audio buzzing artifacts
MDX23C/RoFormer model handling - Some models have tensor alignment issues
Advanced chunked processing with overlap blending for large audio files

📋 Technical Enhancements:

Fixed tensor reshaping and return type compatibility
Enhanced UVR5 model compatibility with warnings
Streamlined audio processing with standardized formats
Improved error handling and user feedback

[4.1.0] - 2025-08-08

Fixed

Add comprehensive RVC (Real-time Voice Conversion) integration

🎵 RVC Voice Conversion:

New Load RVC Character Model node for .pth model loading
RVC Engine support in unified Voice Changer
Iterative refinement passes with smart caching system
Official model download sources (RMVPE, content-vec-best)
Automatic Faiss index loading for enhanced quality
Integration with existing TTS workflow

🔧 Technical Improvements:

Minimal reference wrapper for compatibility
Cache system prevents recomputation of refinement passes
Official download sources from lj1995 and lengyue233
Added required dependencies: faiss-cpu, onnxruntime, torchcrepe

📋 Requirements:

Updated requirements.txt with RVC dependencies
Compatible with existing ChatterBox and F5-TTS workflows

[4.0.0] - 2025-08-06

🚨 BREAKING CHANGES

Complete architectural transformation to TTS Audio Suite
- Project evolved from ChatterBox-focused implementation to universal multi-engine TTS system
- ⚠️ WORKFLOW COMPATIBILITY BROKEN: Existing workflows require migration to new unified node structure
- New project name reflects expanded scope beyond ChatterBox to support multiple TTS engines

Added

🏗️ MAJOR: Unified Multi-Engine Architecture
- Universal TTS nodes that work with any engine (TTS Text, TTS SRT, Voice Changer)
- Engine configuration nodes for easy switching between ChatterBox, F5-TTS, and future engines
- Modular engine adapter system for seamless engine integration
- Character Voices node providing NARRATOR_VOICE outputs for any TTS node
🔧 Advanced Engine Management
- Engine-specific configuration nodes (ChatterBox Engine, F5-TTS Engine)
- Engine adapter pattern for standardized interfaces
- Separation of engine logic from user interface
📁 Complete Project Restructure
- Engine implementations in engines/chatterbox/ and engines/f5tts/
- Unified interface nodes in nodes/unified/
- Engine adapters in engines/adapters/
- Comprehensive utility systems in utils/
- Clear separation between engine-agnostic and engine-specific functionality

Changed

🎯 Node Architecture
- Text and SRT processing now handled by separate, engine-agnostic unified nodes
- Consistent interface across all engines through unified nodes
- Enhanced node categorization for better organization in ComfyUI
⚡ Performance Optimizations
- Cache-aware model loading system prevents unnecessary model reloads
- Smart language grouping processes SRT files by language to reduce model switching overhead
- Optimized memory usage through intelligent model lifecycle management

Technical Details

Engine Adapter Pattern: Standardized interface allowing easy addition of new TTS engines (RVC, Tortoise, etc.)
Unified Caching: Consistent cache management across all engines with engine-specific keys
Modular Design: Clear separation between engine implementations, adapters, and unified interface
Future-Proof Architecture: Foundation for supporting additional TTS engines beyond ChatterBox and F5-TTS

Migration Required

⚠️ Complete workflow migration required - old workflows are incompatible with v4
This is a new project (TTS Audio Suite) separate from the original ChatterBox project
Users must recreate workflows using new unified node structure

This release represents a fundamental architectural transformation, evolving from a ChatterBox extension to a universal multi-engine TTS platform capable of supporting any TTS engine while maintaining the same user experience.

[3.4.3] - 2025-08-05

Fixed

Fix language switching not working properly and add support for flexible language aliases like [German:], [Brazil:], [USA:]

[3.4.2] - 2025-08-05

Fixed

Fix character tag removal bug in single character mode
- Root cause: TTS nodes bypassed character parser in single character mode
- Affected: Both ChatterBox TTS and F5-TTS nodes when text contains unrecognized character tags like [Alex]
- Result: Character tags are now properly removed before TTS generation
- Behavior: Text '[Alex] Hello world' now correctly generates 'Hello world' instead of 'Alex Hello world'

[3.4.1] - 2025-08-03

Changed

🏗️ Major Project Restructure - Complete reorganization for better maintainability
- Engine-centric architecture with separated engines/chatterbox/ and engines/f5tts/
- Organized nodes into nodes/chatterbox/, nodes/f5tts/, nodes/audio/, nodes/base/
- Replaced core/ with organized utils/ structure (audio, text, voice, timing, models, system)
- Self-documenting filenames for better code navigation
- Scalable structure for future engine additions
- All functionality preserved with full backward compatibility
📋 Developer Experience
- Enhanced version bump script with multiline changelog support
- Improved project structure documentation
- Better error handling and import management

[3.4.0] - 2025-08-02

Added

Major Feature: Language Switching with Bracket Syntax
- Introduced [language:character] syntax for inline language switching
- Support for [fr:Alice], [de:Bob], [es:] patterns in text
- Language codes automatically map to appropriate models (F5-DE, F5-FR, German, Norwegian, etc.)
- Character alias system integration with language defaults
- Automatic fallback to English model for unsupported languages with warnings
Language Support
- F5-TTS: English, German (de), Spanish (es), French (fr), Italian (it), Japanese (jp), Thai (th), Portuguese (pt)
- ChatterBox: English, German (de), Norwegian (no/nb/nn)
Modular Architecture
- Modular multilingual engine architecture with engine-specific adapters
- Unified audio cache system with engine-specific cache key generation

Fixed

Fixed character parser regex bug to support empty character names like [fr:]
Character audio tuple handling fixes for ChatterBox engine

Changed

Performance Optimizations
- Smart language loading: SRT nodes now analyze subtitles before model initialization
- Eliminated wasteful default English model loading on startup
- Language groups processed alphabetically (de→en→fr) for predictable behavior
- Reduced model switching overhead in multilingual SRT processing
Technical Improvements
- Enhanced logging to distinguish SRT-level vs multilingual engine operations

[3.3.0] - 2025-08-01

Added

Major Feature: Multilanguage ChatterBox Support
🌍 NEW: Multi-language ChatterBox TTS
Added language parameter as second input in both TTS nodes
All example workflows updated for new parameter structure

Fixed

Language dropdown for English, German, Norwegian models
Automatic HuggingFace model download and management
Local model prioritization for faster generation
Safetensors format support with .pt backward compatibility
Language-aware caching system to prevent model conflicts
ChatterBox TTS Node: Full multilanguage support
ChatterBox SRT TTS Node: SRT timing with multilanguage models
Character switching works seamlessly with all supported languages
Existing workflows need manual parameter adjustment
Robust fallback system: local → HuggingFace → English fallback
JaneDoe84's safetensors loading fix integrated safely
Language-aware cache keys prevent cross-language conflicts

Changed

🎯 Enhanced Nodes:
⚠️ BREAKING CHANGE: Workflow Compatibility
🔧 Technical Improvements:
Enhanced model manager with language-specific loading

[3.2.9] - 2025-08-01

Fixed

Fix seed validation range error - clamp seed values to NumPy valid range (0 to 2^32-1)

[3.2.8] - 2025-07-27

Added

Add graceful fallback when PortAudio is missing
Add startup diagnostic for missing dependencies

Fixed

Fix PortAudio dependency handling for voice recording

Changed

Update README with system dependency requirements

[3.2.7] - 2025-07-23

Fixed

Fix SRT node crash protection template not respecting user input

[3.2.6] - 2025-07-23

Fixed

Fix F5-TTS progress bars and variable scope issues

[3.2.5] - 2025-07-23

Added

Dynamic Model Discovery: Automatically detect local models in ComfyUI/models/F5-TTS/ directory
Multi-Language Support: Added support for 9 language variants (German, Spanish, French, Japanese, Italian, Thai, Brazilian Portuguese)
Custom Download Logic: Implemented language-specific model repository structure handling
Smart Model Config Detection: Automatic model config detection based on folder/model name
Enhanced Model Support: F5-DE, F5-ES, F5-FR, F5-JP, F5-IT, F5-TH, F5-PT-BR alongside standard models

Fixed

Config Mismatch Issues: Resolved configuration problems affecting audio quality
Vocabulary File Handling: Smart handling of vocabulary files for different language models
Cross-Platform Compatibility: Improved international character set support

Changed

Model Loading System: Normalized model loading across base and language-specific models
Error Handling: Enhanced error handling and console output for better debugging
Download Warnings: Added download size and quality warnings for specific models
Model Name Handling: Improved model name handling and caching mechanisms

Technical Note: Resolves GitHub issue #3, significantly improving F5-TTS model detection and language support capabilities.

[3.2.4] - 2025-07-23

Added

Add concatenate timing mode for line-by-line processing without timing constraints
Add concatenate option to timing_mode dropdown in both ChatterBox SRT and F5-TTS SRT nodes
Implement TimingEngine.calculate_concatenation_adjustments() for sequential timing calculations
Add AudioAssemblyEngine.assemble_concatenation() with optional crossfading support
Enhanced reporting system shows original SRT → new timings with duration changes

Fixed

Fastest processing mode with zero audio manipulation for highest quality
Perfect for long-form content while maintaining line-by-line SRT processing benefits

[3.2.3] - 2025-07-22

Added

Add snail 🐌 and rabbit 🐰 emojis to stretch-to-fit timing reports for compress and expand modes in both ChatterBox and F5-TTS SRT nodes

[3.2.2] - 2025-07-21

Added

Add detailed F5-TTS diagnostic messages to help users troubleshoot installation issues. F5-TTS import errors are now always shown during initialization, making it easier to identify missing dependencies without requiring development mode.

[3.2.1] - 2025-07-19

Changed

Voice Conversion Enhancements: Iterative refinement with intelligent caching system for progressive quality improvement and instant experimentation

[3.2.0] - 2025-07-19

Added

MAJOR NEW FEATURES:
Automatic processing with no additional UI parameters
Added full caching support to ChatterBox TTS and F5-TTS nodes
Implemented stable audio component hashing for consistent cache keys
This release brings substantial performance improvements and new creative possibilities for speech generation workflows!

Fixed

Version 3.2.0: Pause Tags System and Universal Caching
⏸️ Pause Tags System - Universal pause insertion with intelligent syntax
Smart pause syntax: [pause:1s], [pause:500ms], [pause:2]
Seamless character integration and parser protection
Universal support across all TTS nodes (ChatterBox, F5-TTS, SRT)
🚀 Universal Audio Caching - Comprehensive caching system for all nodes
Intelligent cache keys prevent invalidation from temporary file paths
Individual segment caching with character-aware separation
Cache hit/miss logging for performance monitoring
🔧 Cache Architecture Overhaul
Fixed cache instability issues across all SRT and TTS nodes
Resolved cache lookup/store mismatch causing permanent cache misses
Optimized pause tag processing to cache text segments independently
Fixed character parser conflicts with pause tag detection
🛠️ Code Quality & Performance
Streamlined codebase with comprehensive pause tag processor

Changed

Intelligent caching: pause changes don't invalidate text cache
Significant speed improvements for iterative workflows
TECHNICAL IMPROVEMENTS:
🎭 Character System Enhancements
Updated text processing order for proper pause/character integration
Enhanced character switching compatibility with pause tags
Improved progress messaging consistency across all nodes
Enhanced crash protection integration with pause tag system

Removed

Removed unnecessary enable_pause_tags UI parameters (automatic now)

[3.1.4] - 2025-07-18

Added

Clean up ChatterBox crash prevention and rename padding parameter

[3.1.3] - 2025-07-18

Fixed

ChatterBox character switching crashes with short text segments by implementing dynamic space padding
Sequential generation CUDA tensor indexing errors in character switching mode
Version bump script now prevents downgrade attempts

[3.1.2] - 2025-07-17

Added

Implement user-friendly character alias system with #character_alias_map.txt file
Add comprehensive alias documentation to CHARACTER_SWITCHING_GUIDE.md with examples
Update README features to highlight new alias system and improve emoji clarity

Fixed

Support flexible alias formats: 'Alias = Character' and 'Alias[TAB]Character' with smart parsing
Replace old JSON character_alias_map.json with more accessible text format
Maintain backward compatibility with existing JSON files for seamless migration

[3.1.1] - 2025-07-17

Added

Update character switching documentation to reflect new system

Fixed

Fix character discovery system to use filename-based character names instead of folder names
Folders now used for organization only, improving usability and clarity

[3.1.0] - 2025-07-17

Added

🎭 Character Switching System

NEW: Universal [Character] tag support across all TTS nodes
NEW: Character alias mapping with JSON configuration files
NEW: Dual voice discovery (models/voices + voices_examples directories)
NEW: Line-by-line character parsing for natural narrator fallback
NEW: Robust fallback system for missing characters
ENHANCED: Voice discovery with flat file and folder structure support
ENHANCED: Character-aware caching system
DOCS: Added comprehensive CHARACTER_SWITCHING_GUIDE.md

🎙️ Overlapping Subtitles Support

NEW: Support for overlapping subtitles in SRT nodes
NEW: Automatic mode switching (smart_natural → pad_with_silence)
NEW: Enhanced audio mixing for conversation patterns
ENHANCED: SRT parser with overlap detection and optional validation
ENHANCED: Audio assembly with overlap-aware timing

Enhanced

🔧 Technical Improvements

ENHANCED: SRT parser preserves newlines for character switching
ENHANCED: Character parsing with punctuation normalization
ENHANCED: Voice discovery initialization on startup
ENHANCED: Timing reports distinguish original vs generated overlaps
ENHANCED: Mode switching info displayed in generation output

Fixed

FIXED: Line-by-line processing in SRT mode for proper narrator fallback
FIXED: Character tag removal before TTS generation
FIXED: "Back to me" bug in character parsing
FIXED: ChatterBox SRT caching issue with character system
FIXED: UnboundLocalError in timing mode processing

[3.0.13] - 2025-07-16

Added

Add F5-TTS SRT workflow and fix README workflow links
Added new F5-TTS SRT and Normal Generation workflow

Fixed

Fixed broken SRT workflow link in README (missing emoji prefix)
All workflow links now point to correct files

Changed

Updated workflow section to properly categorize Advanced workflows

[3.0.12] - 2025-07-16

Added

Added F5-TTS availability checking to initialization messages

Fixed

Fix F5-TTS model switching and improve initialization messages
Fixed F5-TTS model cache not reloading when changing model names
Removed redundant SRT success messages (only show on actual issues)
Enhanced error handling for missing F5-TTS dependencies

Changed

Improved F5-TTS model loading to only check matching local folders

[3.0.11] - 2025-07-16

Removed

Optimize dependencies - remove unused packages to reduce installation time and conflicts

[3.0.10] - 2025-07-15

Fixed

Fix missing diffusers dependency
Fix record button not showing due to node name mismatch in JavaScript extension

[3.0.9] - 2025-07-15

Added

Add enhanced voice discovery system with dual folder support for F5-TTS nodes

[3.0.8] - 2025-07-15

Fixed

Fix tensor dimension mismatch in audio concatenation for 5+ TTS chunks

[3.0.7] - 2025-07-15

Added

Add comprehensive parameter migration checklist documentation

Fixed

Improve F5-TTS Edit parameter organization and fix RMS normalization
Move target_rms to advanced options as post_rms_normalization for clarity
Fix RMS normalization to preserve original segments volume

Removed

Remove non-functional speed parameter from edit mode

[3.0.6] - 2025-07-15

Fixed

Fix SRT package naming conflict - resolves issue #2
Rename internal 'srt' package to 'chatterbox_srt' to avoid conflict with PyPI srt library

Changed

Update all imports in nodes/srt_tts_node.py and nodes/f5tts_srt_node.py

[3.0.5] - 2025-07-14

Fixed

Fix import detection initialization order - resolves ChatterboxTTS availability detection

[3.0.4] - 2025-07-14

Fixed

Fix ChatterBox import detection to find bundled packages

[3.0.3] - 2025-07-14

Fixed

Fix F5-TTS device mismatch error for MP3 audio editing

[3.0.2] - 2025-07-14

Added

Fix Features section formatting for proper GitHub markdown rendering
Add placeholder for F5-TTS Audio Analyzer screenshot
Restructure SRT_IMPLEMENTATION.md documentation
Add comprehensive table of contents and Quick Start section
Add enhanced version bumping scripts with multiline support
Create automated changelog generation with categorization

Fixed

Documentation improvements and fixes
Fix README.md formatting and image placement
Restore both ChatterBox TTS and Voice Capture images side by side
Fix code block formatting and usage examples for ComfyUI users
Polish language and maintain professional tone
Version management automation system
Optimize CLAUDE.md for token efficiency

Changed

Improve image organization and section clarity
Improve document organization with Quick Reference tables

[3.0.1] - 2025-07-14

Fixed

Added sounddevice to requirements.txt to prevent ModuleNotFoundError when using voice recording functionality
Removed optional sounddevice installation section from README as it's now included by default

Changed

Voice recording dependencies are now installed automatically with the main requirements
Simplified installation process by removing optional dependency steps

[3.0.0] - 2025-07-13

Added

Implemented F5-TTS nodes for high-quality voice synthesis with reference audio + text cloning.
Added Audio Wave Analyzer node for interactive waveform visualization and precise timing extraction for F5-TTS workflows. 📖 Complete Guide
Added F5TTSEditNode for speech editing workflows.
Added F5TTSSRTNode for generating TTS from SRT files using F5-TTS.

New Nodes

F5TTSNode
F5TTSSRTNode
F5TTSEditNode
AudioAnalyzerNode
AudioAnalyzerOptionsNode

Contributors

Diogod

[2.0.2] - 2025-06-27

Fixed

Transformers Compatibility: Fixed compatibility issues with newer versions of the transformers library after ComfyUI updates
- Resolved LlamaModel.__init__() got an unexpected keyword argument 'attn_implementation' error by removing direct parameter passing to LlamaModel constructor
- Fixed PretrainedConfig.update() got an unexpected keyword argument 'output_attentions' error by using direct attribute setting instead of config.update()
- Fixed DynamicCache.update() missing 2 required positional arguments error by simplifying cache handling to work with different transformers versions
Cache Management: Updated cache handling in the T3 inference backend to be compatible with both older and newer transformers API versions
Configuration Safety: Added safer configuration handling to prevent compatibility issues across different transformers versions

Improved

Error Reporting: Enhanced error messages in model loading to provide better debugging information
Version Compatibility: Made the codebase more resilient to transformers library version changes

[2.0.1] - 2025-06-17

Changed

Node Renaming for Conflict Resolution: Renamed nodes to avoid conflicts with the original ChatterBox Voice repository
Added "(diogod)" suffix to distinguish from original implementation:
- ChatterBoxVoiceTTS → ChatterBoxVoiceTTSDiogod (displayed as "🎤 ChatterBox Voice TTS (diogod)")
- ChatterBoxVoiceVC → ChatterBoxVoiceVCDiogod (displayed as "🔄 ChatterBox Voice Conversion (diogod)")
- ChatterBoxVoiceCapture → ChatterBoxVoiceCaptureDiogod (displayed as "🎙️ ChatterBox Voice Capture (diogod)")
Note: "📺 ChatterBox SRT Voice TTS" remains unchanged as it was unique to this implementation

[2.0.0] - 2025-06-14

Changed

MAJOR ARCHITECTURAL REFACTORING: Transformed the project from a monolithic structure to a clean, modular architecture
Decomposed the massive 1,922-line nodes.py into specialized, focused modules for improved maintainability and LLM-friendly file sizes
Created structured directory architecture:
- nodes/ - Individual node implementations (tts_node.py, vc_node.py, srt_tts_node.py, audio_recorder_node.py)
- core/ - Core functionality modules (model_manager.py, audio_processing.py, text_chunking.py, import_manager.py)
- srt/ - SRT-specific functionality (timing_engine.py, audio_assembly.py, reporting.py)
Extracted specialized functionality into focused modules:
- Model management and loading logic → core/model_manager.py
- Audio processing utilities → core/audio_processing.py
- Text chunking algorithms → core/text_chunking.py
- Import and dependency management → core/import_manager.py
- SRT timing calculations → srt/timing_engine.py
- Audio segment assembly → srt/audio_assembly.py
- Timing report generation → srt/reporting.py
Integrated audio recorder node functionality into the unified architecture
Established clean separation of concerns with well-defined interfaces between modules
Implemented proper inheritance hierarchy with BaseNode class for shared functionality

Fixed

Resolved original functionality issues discovered during the refactoring process
Fixed module import paths and dependencies across the codebase
Corrected audio processing pipeline inconsistencies
Addressed timing calculation edge cases in SRT generation

Maintained

100% backward compatibility - All existing workflows and integrations continue to work without modification
Preserved all original API interfaces and node signatures
Maintained feature parity across all TTS, voice conversion, and SRT generation capabilities
Kept all existing configuration options and parameters intact

Improved

Maintainability: Each module now has a single, well-defined responsibility
Readability: Code is organized into logical, easily navigable modules
Testability: Modular structure enables isolated unit testing of individual components
Extensibility: Clean architecture makes it easier to add new features and nodes
LLM-friendly: Smaller, focused files are more manageable for AI-assisted development
Development workflow: Reduced cognitive load when working on specific functionality

Technical Details

Maintained centralized node registration through __init__.py
Preserved ComfyUI integration patterns and node lifecycle management
Kept all original error handling and progress reporting mechanisms
Maintained thread safety and resource management practices

[1.2.0] - 2025-06-13

Updated

Updated README.md and requirements.txt with proactive advice to upgrade pip, setuptools, and wheel before installing dependencies. This aims to prevent common installation issues with s3tokenizer on certain Python environments (e.g., Python 3.10, Stability Matrix setups).

Added

Added progress indicators to TTS generation showing current segment/chunk progress (e.g., "🎤 Generating TTS chunk 2/5..." or "📺 Generating SRT segment 3/124...") to help users estimate remaining time and track generation progress.

Fixed

Fixed interruption handling in ChatterBox TTS and SRT nodes by using ComfyUI's comfy.model_management.interrupt_processing instead of the deprecated execution.interrupt_processing attribute. This resolves the "ComfyUI's 'execution.interrupt_processing' attribute not found" warning and enables proper interruption between chunks/segments during generation.
Fixed interruption behavior to properly signal to ComfyUI that generation was interrupted by raising InterruptedError instead of gracefully continuing. This prevents ComfyUI from caching interrupted results and ensures the node will re-run properly on subsequent executions.
Fixed IndexError crashes in timing report and SRT string generation functions when called with empty lists, adding proper edge case handling for immediate interruption scenarios.

Improved

Improved smart natural timing mode to distinguish between significant and insignificant audio truncations. Truncations smaller than 50ms are now shown as "Fine-tuning audio duration" without the alarming 🚧 emoji, while only meaningful truncations (>50ms) that indicate real timing conflicts are highlighted with the warning emoji. This reduces noise in timing reports and helps users focus on actual issues.
Reduced console verbosity by removing detailed FFmpeg processing messages (filter chains, channel processing details, etc.) during time stretching operations. The timing information is still available in the detailed timing reports, making the console output much cleaner while maintaining full functionality.
Optimized progress messages for SRT generation to only show "📺 Generating SRT segment..." when actually generating new audio, not when loading from cache. This eliminates console spam when cached segments load instantly and provides more accurate progress indication for actual generation work.

Fixed

Fixed sequence numbering preservation in timing reports and Adjusted_SRT output for stretch_to_fit and pad_with_silence modes. All timing modes now correctly preserve the original SRT sequence numbers (e.g., 1, 1, 14) instead of renumbering them sequentially (1, 2, 3), maintaining consistency with smart_natural mode and ensuring more accurate SRT output.

[1.1.1] - 2025-06-11

Fixed

Resolved a tensor device mismatch error (cuda:0 vs cpu) in the "ChatterBox SRT Voice TTS" node. This issue occurred when processing SRT files, particularly those with empty text entries, in "stretch_to_fit" and "pad_with_silence" timing modes. The fix ensures all audio tensors are consistently handled on the target processing device (self.device) throughout the audio generation and assembly pipeline.

[1.1.0] - 2025-06-10

Added

Added the ability to handle subtitles with empty strings or silence in the SRT node.

194 KiB Raw Permalink Blame History

Changelog

[5.8.1] - 2026-08-11

Added

Changed

[5.8.0] - 2026-08-11

Added

Fixed

[5.7.0] - 2026-08-10

Added

Changed

[5.6.5] - 2026-08-03

Fixed

[5.6.4] - 2026-08-03

Added

[5.6.3] - 2026-08-01

Changed

Fixed

[5.6.2] - 2026-07-30

Changed

Fixed

[5.6.1] - 2026-07-30

Fixed

[5.6.0] - 2026-07-25

Added

Changed

[5.5.3] - 2026-07-24

Added

Changed

[5.5.2] - 2026-07-21

Fixed

[5.5.1] - 2026-07-18

Added

Changed

Fixed

[5.5.0] - 2026-07-17

Added

Changed

Removed

[5.4.16] - 2026-07-16

Fixed

[5.4.15] - 2026-07-15

Fixed

[5.4.14] - 2026-07-15

Fixed

[5.4.13] - 2026-07-14

Added

Fixed

[5.4.12] - 2026-07-14

Added

Fixed

[5.4.11] - 2026-07-13

Fixed

[5.4.10] - 2026-07-13

Added

[5.4.9] - 2026-07-13

Added

Changed

Fixed

[5.4.8] - 2026-07-13

Added

Fixed

[5.4.7] - 2026-07-13

Added

Changed

Fixed

[5.4.6] - 2026-07-13

Added

Changed

[5.4.5] - 2026-07-12

Added

Fixed

[5.4.4] - 2026-07-11

Added

Changed

Fixed

[5.4.3] - 2026-07-10

Added

Fixed

[5.4.2] - 2026-07-10

194 KiB

Raw Permalink Blame History