Qwen-ASR Local Speech Recognition Model

What is this?

Qwen-ASR is a local speech recognition model developed by the Alibaba Tongyi Qianwen team. It converts audio speech into text without requiring an internet connection. The model runs entirely locally, protecting your privacy while providing high-accuracy speech recognition.

Select Qwen-ASR (Local) in pyVideoTrans's speech recognition service to use this model.

Version Requirements

Must upgrade to v3.97+ to use the Qwen-ASR local model.

Model Selection

Qwen-ASR offers two model sizes, each with its own trade-offs:

Model	Parameters	Accuracy	Resource Usage	Recommended For
0.6B	600M	High	Lower	Limited VRAM, speed priority
1.7B	1.7B	Higher	Higher	Accuracy priority, sufficient VRAM

• 0.6B model: Smaller, faster inference, lower VRAM requirements. Suitable for average computers.
• 1.7B model: Higher recognition accuracy, but requires more VRAM and compute. Suitable for best-quality seekers.

Model Download

Auto-download (Recommended)

On first use, the software will automatically download the required model from the model repository. The Chinese UI defaults to using the ModelScope domestic mirror.

Manual Download

If auto-download fails or is too slow, download manually.

Step 1: Verify Directory Structure

Ensure the following model directories exist under the software's models/ folder (create them if missing):

app_dir/
└── models/
    ├── models--Qwen--Qwen3-ASR-0.6B/    (0.6B model directory)
    └── models--Qwen--Qwen3-ASR-1.7B/    (1.7B model directory)

Step 2: Download 1.7B Model

1. Open HuggingFace page: https://huggingface.co/Qwen/Qwen3-ASR-1.7B/tree/main
2. Download all files from the page.
3. Place all files in models/models--Qwen--Qwen3-ASR-1.7B/.

Step 3: Download 0.6B Model

1. Open HuggingFace page: https://huggingface.co/Qwen/Qwen3-ASR-0.6B/tree/main
2. Download all files from the page.
3. Place all files in models/models--Qwen--Qwen3-ASR-0.6B/.

Step 4: Select Model

In the software's speech recognition settings, select either 0.6B or 1.7B as needed.

How It Works

Qwen-ASR uses VAD (Voice Activity Detection) for audio preprocessing:

1. Uses the ten-vad model to intelligently segment audio into short clips.
2. Processes clips in batches of 8 for inference.
3. Merges recognition results from all clips.

This approach has low VRAM usage and fast inference, making it suitable for long audio files.

---

FAQ

Issue	Possible Cause	Solution
Model download failed	Network issue or HuggingFace blocked	Use ModelScope mirror or download manually
Empty recognition results	Unsupported audio format	Use common format (wav/mp3/aac), preferably WAV
Out of VRAM	Insufficient GPU memory	Use 0.6B model or close other VRAM-intensive programs
Very slow recognition	GPU acceleration not used	Check CUDA environment configuration
Software freezes	Long audio processing	This is normal — be patient, or use shorter audio