Comfyui_SynVow_Qwen3ASR

<img width="1957" height="648" alt="image" src="https://github.com/user-attachments/assets/b0dfdf1c-1bf2-4f97-a16b-a9615c9f4a4f" /> # ComfyUI SynVow Qwen3-ASR A ComfyUI speech recognition plugin based on [Qwen3-ASR](https://github.com/QwenLM/Qwen3-ASR).

Original Project

https://github.com/QwenLM/Qwen3-ASR

Features

• � Speech-to-Text: Transcribe audio to text with high accuracy
• 🌍 Multi-language Support: Supports 52 languages/dialects with automatic language detection
• ⏱️ Forced Alignment: Generate word/character-level timestamps
• 🤖 Auto Model Download: Automatically download models from HuggingFace on first use
• 📊 Long Audio Support: ASR supports up to 20 minutes, alignment supports up to 3 minutes (auto-chunking for longer audio)

Installation

1. Install Dependencies

pip install -r requirements.txt

Or manually install:

pip install qwen-asr huggingface_hub torchaudio

2. Model Download

Models will be automatically downloaded to the following directory on first use:

ComfyUI/models/Qwen3-ASR/
├── Qwen3-ASR-1.7B/           # ASR model (1.7B, default)
├── Qwen3-ASR-0.6B/           # ASR model (0.6B, lighter)
└── Qwen3-ForcedAligner-0.6B/ # Forced alignment model

Node Description

1. Qwen3-ASR Loader

Load Qwen3-ASR speech recognition model.

Input Parameters:

• model_name: Select model version
  • Qwen3-ASR-1.7B (default) - Better accuracy
  • Qwen3-ASR-0.6B - Faster, lower VRAM

Output:

• model: ASR model for transcription node

2. Qwen3-ASR Transcribe

Transcribe audio to text.

Input Parameters:

• model: Model from Loader node
• audio: Audio input (ComfyUI AUDIO type)
• language: Language selection (default: Auto)

Output:

• text: Transcribed text
• language: Detected language

3. Qwen3 ForcedAligner Loader

Load forced alignment model for generating timestamps.

Input Parameters:

• model_name: Select model version
  • Qwen3-ForcedAligner-0.6B (default)

Output:

• aligner: Aligner model for alignment node

4. Qwen3 Forced Align

Generate word/character-level timestamps.

Input Parameters:

• aligner: Aligner from Loader node
• audio: Audio input (ComfyUI AUDIO type)
• text: Text to align
• language: Language (supports 11 languages)
• segment_by_sentence: Segment by sentence (default: true). When enabled, splits output by punctuation; when disabled, outputs per character/word.

Output:

• timestamps: Timestamps in format text\tstart_time\tend_time
• text_list: Text segments, one per line
• start_times: Start times, one per line
• end_times: End times, one per line

Usage Example

Basic Speech-to-Text

1. Use Load Audio node to load audio file
2. Connect Qwen3-ASR Loader to load model
3. Connect Qwen3-ASR Transcribe for transcription
4. Output text and detected language

With Timestamps

1. Load Audio → Qwen3-ASR Loader → Qwen3-ASR Transcribe → Get text
2. Qwen3 ForcedAligner Loader → Qwen3 Forced Align (input audio + text) → Get timestamps

Supported Languages

ASR (52 languages): Chinese, English, Cantonese, Arabic, German, French, Spanish, Portuguese, Indonesian, Italian, Korean, Russian, Thai, Vietnamese, Japanese, Turkish, Hindi, Malay, Dutch, Swedish, Danish, Finnish, Polish, Czech, Filipino, Persian, Greek, Romanian, Hungarian, Macedonian, etc.

Forced Alignment (11 languages): Chinese, English, Cantonese, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish

Notes

1. VRAM Requirements:

   • Qwen3-ASR-0.6B: ~4GB VRAM
   • Qwen3-ASR-1.7B: ~8GB VRAM
   • ForcedAligner-0.6B: ~4GB VRAM

2. Audio Length Limits:

   • ASR: Max 20 minutes (1200 seconds)
   • Forced Alignment: Max 3 minutes (180 seconds)
   • Longer audio is automatically chunked and merged

License

This project follows the license of the original Qwen3-ASR project.

Related Links

• Qwen3-ASR Original Project
• Qwen3-ASR Models on HuggingFace

Changelog

2026-02-06

• Qwen3 Forced Align: Added segment_by_sentence parameter (boolean, default: true) to split alignment output by sentence/phrase based on punctuation in input text
• Qwen3 Forced Align: Added 3 new outputs: text_list, start_times, end_times for easier downstream processing