1038lab-ComfyUI-QwenASR/Update.md

Update v1.1.0

This release focuses on delivering production-ready speech transcription and subtitling in ComfyUI, improving output accuracy, and expanding the node suite into a complete audio workflow.

---

1. How v1.1.0 Enriches the Node Family

Instead of relying on a single speech-to-text node, users now have a complete three-node toolkit tailored to different audio tasks:

• ASR (QwenASR): Fast, lightweight speech transcription for voice prompts, dialogue notes, and quick dictation.
• Subtitle (QwenASR): Automatically chunks speech into natural, timestamped subtitle sentences with one-click export to .srt or .txt.
• Forced Align (QwenASR): High-precision word-level aligner built specifically to process long continuous speech (podcasts, lectures, long videos) without hitting duration limits.

2. How Transcription Output is Made Significantly More Accurate

Raw speech recognition models output verbatim acoustic sounds, which often produces awkward, unreadable results. Version 1.1.0 introduces smart Inverse Text Normalization (ITN):

• Clean Numerical Formatting: Spoken numbers, decimals, and percentages are automatically formatted into standard digits (e.g. 一百二十八 → 128, 两千 → 2000, 一点七 → 1.7, 百分之九十九 → 99%).
• Acronym Clean-Up: Spaced-out letters from phonetic transcription are merged into clean abbreviations (e.g. A S R → ASR, U S B → USB, A P I → API).
• Multi-Language Phonetic & Term Corrections: Built-in rules correct common spoken tech and AI terminology across English, Chinese, Japanese, Korean, and French (e.g. ChatGPT, OpenAI, DeepSeek, ComfyUI).
• Idiom & Phrase Safeguards: Built-in whitelist protects idioms from being accidentally corrupted into digits (e.g. 万无一失, 五颜六色).

3. Major Architectural Upgrade: Official Native Models & Transformers 5

This version introduces a significant architectural evolution for stability and performance:

• Full Migration to Official Hugging Face Native Models: Transitioned completely to official native checkpoints (Qwen3-ASR-1.7B-hf, Qwen3-ASR-0.6B-hf, Qwen3-ForcedAligner-0.6B-hf). Legacy un-suffixed checkpoints are now deprecated in favor of official standard pipelines.
• Upgraded to transformers >= 5.13.0: Upstream Transformers 5 integration eliminates bulky, brittle legacy backend code, providing pure PyTorch multimodal execution, lower VRAM usage, and significantly faster inference across all hardware.

4. New Features to Help Users Get Exactly What They Need

• One-Click Transcription & Alignment: In the Forced Align node, leaving the transcript text empty automatically transcribes the audio and produces precision word timestamps in a single step.
• Zero-Restart Custom Dictionary (itn_rules.json): Add custom industry terms, names, or words in your own language into itn_rules.json. Changes apply immediately on the next generation without needing to restart ComfyUI.
• Apple Silicon Mac Support: Automatic precision safeguards (fp16 fallback and SDPA attention) provide smooth, crash-free execution on M-series Macs.
• Dual High-Speed Download Mirrors via config.json: Set your preferred download mirror (HuggingFace or ModelScope) in config.json for fast, reliable model downloads anywhere in the world.