[Support] Qwen3-ASR-API - Pure OpenAI-compatible Speech-to-Text API powered by Qwen3-ASR

What it is

Pure OpenAI-compatible Speech-to-Text API powered by Qwen3-ASR (Alibaba Qwen). State-of-the-art accuracy for Chinese (22 dialects), English, and 50+ languages. No bloat — just the model served via vLLM.

Exposes /v1/audio/transcriptions (OpenAI-compatible), /docs (Swagger), /health

Template details

• App name: Qwen3-ASR-API

• Template URL: https://raw.githubusercontent.com/hsiang-han/unraid_templates/main/templates/qwen3-asr-api.xml

• Icon URL: https://raw.githubusercontent.com/hsiang-han/unraid_templates/main/assets/qwen3-asr-api-icon.png

• Project URL: https://github.com/hsiang-han/qwen3-asr-api

• Support URL: https://github.com/hsiang-han/qwen3-asr-api/issues

Container image strategy

• Repository (GPU only): ghcr.io/hsiang-han/qwen3-asr-api:latest

• Extra Parameters: --gpus all --shm-size=4g

• CUDA 12.8 — supports Blackwell (RTX 5060/5070/5090) and Ada GPUs

• Requires NVIDIA driver >= 570

Model selection

Switch via MODEL_ID environment variable (restart required):

• Qwen/Qwen3-ASR-0.6B (default) — ~2-3GB VRAM, RTFx 166, low latency

• Qwen/Qwen3-ASR-1.7B — ~4-6GB VRAM, RTFx 148, best accuracy

Default mappings / settings

• Port: 8000

• Host path: /mnt/user/appdata/qwen3-asr-api/models → /root/.cache/huggingface

• API docs: http://[IP]:[PORT:8000]/docs

• First start downloads model (~1-3GB), subsequent starts are fast

Environment variables

• MODEL_ID = Qwen/Qwen3-ASR-0.6B — Model to serve

• GPU_MEMORY_UTILIZATION = 0.8 — GPU memory fraction (0.0-1.0)

• NVIDIA_VISIBLE_DEVICES = all — GPU selection

• MAX_MODEL_LEN = 8192 — Max sequence length for KV cache. Default value can supports ~10 min audio. Lower to save VRAM, raise for longer audio.

Usage

curl -X POST http://YOUR-UNRAID-IP:8000/v1/audio/transcriptions -F "[email protected]" -F "model=qwen3-asr"

Works with any OpenAI-compatible client.

Edited June 8Jun 8 by hsiang