Qwen ASR Skill - Local Speech-to-Text
Transcribe audio files to text locally using the Qwen3-ASR-1.7B model.
Setup
Install dependencies into a Python 3.10+ virtual environment:
python -m venv .venv
source .venv/bin/activate
pip install qwen-asr soundfile silero-vad
For GPUs with compute capability < 7.0 (e.g. GTX 1060), install PyTorch 2.4.x with CUDA 11.8:
pip install torch==2.4.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu118
Usage
Run the transcription script with the path to an audio file:
python scripts/transcribe.py <audio_path>
Parameters
| Parameter | Description | Default |
|---|---|---|
audio_path | Absolute path to audio file (required) | - |
--language | Force language (e.g. Chinese, English). Auto-detect if omitted | Auto-detect |
--device | Inference device: auto / cuda / cpu | auto |
--model-path | Model path or HuggingFace ID | ~/models/Qwen3-ASR-1.7B |
--max-chunk-sec | Max chunk duration for VAD splitting. Long audio is split at silence boundaries | 90 |
--max-new-tokens | Max tokens to generate. Increase for long audio | 2048 |
Examples
Basic transcription:
python scripts/transcribe.py /path/to/audio.wav
Force language:
python scripts/transcribe.py /path/to/audio.mp3 --language Chinese
Force CPU inference:
python scripts/transcribe.py /path/to/audio.flac --device cpu
Output Format
The script outputs JSON to stdout and status info to stderr:
{"language": "Chinese", "text": "Transcribed text content"}
On error:
{"error": "Error description"}
Notes
- First run downloads the model (~4.7GB), cached for subsequent runs
- Auto mode: tries GPU (float16) first, falls back to CPU (float32) if VRAM is insufficient
- Supports: WAV, MP3, FLAC, M4A, OGG and other common audio formats
- 52 languages including Chinese, English, Japanese, Korean, French, German, etc.
- 22 Chinese dialects supported
- Long audio: Audio longer than 90s is automatically split at silence boundaries using silero-vad, transcribed chunk by chunk, then concatenated. This prevents OOM on limited VRAM GPUs.