ComfyUI-Qwen3-ASR

ComfyUI-Qwen3-ASR

ComfyUI custom nodes for Qwen3-ASR transcription.

This package wraps the official qwen-asr Python package and defaults to Qwen/Qwen3-ASR-0.6B for practical local testing. Qwen/Qwen3-ASR-1.7B is also available from the model selector.

Nodes

• Qwen3 ASR - Transcribe File
  • Selects an audio/video file from ComfyUI's input directory.
• Qwen3 ASR - Transcribe Audio
  • Accepts a standard ComfyUI AUDIO input.

Use the model_name dropdown for known models, or set custom_model_id to a new Hugging Face model id when Qwen publishes a newer qwen-asr compatible variant. When custom_model_id is not empty, it takes precedence over model_name.

Both nodes output:

• transcript: recognized speech text
• language: detected or requested language
• metadata_json: JSON metadata returned by the model wrapper

Word-level timestamps are not exposed in this initial ASR node because qwen-asr requires the separate Qwen3 ForcedAligner model for that mode.

Install

Install dependencies in the same Python environment used by ComfyUI:

pip install -r ComfyUI/custom_nodes/ComfyUI-Qwen3-ASR/requirements.txt

Restart ComfyUI after installation.

The node installs qwen-asr from the compatibility fork:

git+https://github.com/endman100/Qwen3-ASR.git#egg=qwen-asr

The fork intentionally uses lower-bound-only runtime dependencies so newer Qwen3-ASR runtime/model support can be picked up with:

pip install -U -r ComfyUI/custom_nodes/ComfyUI-Qwen3-ASR/requirements.txt

When newer dependency versions break compatibility, fix the fork and reinstall from this node's requirements.txt.

Notes

The first run downloads model weights through Hugging Face. Qwen3-ASR accepts full language names such as English and Chinese; this node maps common ISO codes such as en, zh, and ja to the names expected by qwen-asr.

Sources:

• https://huggingface.co/Qwen/Qwen3-ASR-0.6B
• https://huggingface.co/Qwen/Qwen3-ASR-1.7B
• https://huggingface.co/collections/Qwen/qwen3-asr