Qwen3-ASR Docker

English | 简体中文 | 繁體中文 | 日本語

🎙️ Qwen3-ASR Docker Deployment

Docker Pulls GitHub Stars License

Production-ready All-in-One Docker deployment for Qwen3-ASR — Alibaba's state-of-the-art open-source ASR supporting 52 languages/dialects, timestamps, streaming, dark-theme UI, REST API & MCP.

---

✨ Features

• 🌐 52 Languages & Dialects — 30 languages + 22 Chinese dialects, auto language detection
• ⏱️ Forced Alignment Timestamps — word/character level via Qwen3-ForcedAligner-0.6B
• 🔄 Streaming Transcription — real-time results via WebSocket
• 🎨 Dark Theme UI — glassmorphism design, 4-language i18n (EN/CN/TW/JP), mic recording + file upload
• 🚀 GPU Management — lazy loading, auto-offload after idle, manual release, model switching
• 📡 REST API — FastAPI with Swagger docs, OpenAPI schema
• 🔌 MCP Integration — fastmcp server with 4 tools for AI agent integration
• 🐳 All-in-One Docker — 3 models embedded, zero runtime download
• 📊 Two Models — Qwen3-ASR-1.7B (highest accuracy) and Qwen3-ASR-0.6B (fast & efficient)

---

🚀 Quick Start

Docker One-Liner

docker run -d --gpus '"device=2"' --name qwen3-asr \
  -p 8250:8200 -p 8251:8201 \
  --restart unless-stopped \
  neosun/qwen3-asr:latest

Open http://localhost:8250 for UI, http://localhost:8250/docs for API docs.

Docker Compose

git clone https://github.com/neosun100/Qwen3-ASR.git
cd Qwen3-ASR
bash start.sh  # auto-selects freest GPU

---

📡 API Reference

Endpoint	Method	Description
/	GET	Web UI
/health	GET	Health check (version, GPU, model status)
/api/status	GET	Detailed status (GPU, models, supported languages)
/api/transcribe	POST	Speech recognition (multipart/form-data)
/api/transcribe/stream	WebSocket	Streaming transcription
/api/languages	GET	Supported languages list
/api/gpu-offload	POST	Release GPU memory
/docs	GET	Swagger API documentation

Transcribe Parameters

Parameter	Type	Default	Description
file	file	required	Audio file (WAV/MP3/FLAC/M4A/OGG)
language	string	auto	Language name or auto
model	string	Qwen3-ASR-1.7B	Qwen3-ASR-1.7B or Qwen3-ASR-0.6B
return_timestamps	bool	false	Return word/char timestamps
dtype	string	bfloat16	bfloat16 or float16

Example

curl -X POST http://localhost:8250/api/transcribe \
  -F '[email protected]' \
  -F 'language=auto' \
  -F 'model=Qwen3-ASR-1.7B' \
  -F 'return_timestamps=true'

{
  "text": "Hello world",
  "language": "English",
  "timestamps": [
    {"text": "Hello", "start": 0.0, "end": 0.5},
    {"text": "world", "start": 0.5, "end": 1.0}
  ],
  "duration_seconds": 1.0,
  "process_time_seconds": 0.15,
  "rtf": 0.15
}

---

🔌 MCP Integration

MCP server runs on port 8251. Config for Claude Desktop / Cursor / Kiro:

{
  "mcpServers": {
    "qwen3-asr": {
      "command": "python",
      "args": ["app/mcp_server.py"],
      "env": {
        "MODEL_PATH_QWEN3_ASR_1_7B": "/models/Qwen3-ASR-1.7B"
      }
    }
  }
}

Available Tools: transcribe, get_status, get_languages, gpu_offload

---

🏗️ Tech Stack

Component	Technology
ASR Engine	Qwen3-ASR (0.6B / 1.7B)
Forced Aligner	Qwen3-ForcedAligner-0.6B
Backend	FastAPI + Uvicorn
Frontend	Pure HTML/CSS/JS (no framework)
MCP Server	fastmcp
Container	NVIDIA CUDA 12.4 + Ubuntu 22.04
GPU Mgmt	Auto-offload, lazy loading

---

📁 Project Structure

app/
├── server.py              # FastAPI backend
├── gpu_manager.py         # GPU resource management
├── mcp_server.py          # MCP server (fastmcp)
└── templates/
    └── index.html         # Dark theme UI
tests/
├── test_api.py            # 22 API tests
└── test_mcp.py            # 8 MCP tests
Dockerfile                 # All-in-One image
docker-compose.yml         # GPU + health check
start.sh                   # One-click launcher

---

⚙️ Configuration

Variable	Default	Description
GPU_ID	2	GPU device ID
PORT	8200	API server port
MCP_PORT	8201	MCP server port
GPU_IDLE_TIMEOUT	600	Auto-offload timeout (seconds)

Copy .env.example to .env and edit as needed.

---

🌐 Online Demo

https://qwen3-asr.aws.xin

---

📄 License

Apache-2.0. Based on Qwen3-ASR by Alibaba Qwen Team.

---

⭐ Star History

Star History Chart