Qwen3-ASR-API

English | 中文

Pure OpenAI-compatible Speech-to-Text API powered by Qwen3-ASR.

No extra services, no NGINX, no voiceprint database — just the model served via vLLM with an OpenAI-compatible endpoint.

What this adds

The official qwenllm/qwen3-asr Docker image has no entrypoint (drops to interactive shell), making it unusable on platforms like Unraid. This project adds an entrypoint for out-of-the-box usage, compatible with any Docker environment:

• Auto-start qwen-asr-serve on container launch
• Environment variable for model switching (no rebuild needed)
• GPU memory control via env var
• Unraid Community Applications template

Quick Start

docker run -d --gpus all --shm-size=4g \
  -p 8000:80 \
  -v /path/to/models:/root/.cache/huggingface \
  -e MODEL_ID=Qwen/Qwen3-ASR-0.6B \
  ghcr.io/hsiang-han/qwen3-asr-api:latest

First start downloads the model (~1-3GB depending on variant).

Usage (OpenAI-compatible)

curl -X POST http://localhost:8000/v1/audio/transcriptions \
  -H "Content-Type: multipart/form-data" \
  -F "file=@audio.wav" \
  -F "model=qwen3-asr"

Or with OpenAI SDK:

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="none")
result = client.audio.transcriptions.create(
    model="qwen3-asr",
    file=open("audio.wav", "rb")
)
print(result.text)

Model Options

Switch by changing MODEL_ID env var and restarting container.

Unraid Install

1. Add template repo: https://github.com/hsiang-han/unraid_templates
2. Find "Qwen3-ASR-API" in Community Applications
3. Configure MODEL_ID and GPU settings
4. Start — first launch downloads model, subsequent starts are fast

Environment Variables

License

Apache-2.0 (same as upstream Qwen3-ASR)