Qwen3-TTS-API

中文文档

OpenAI-compatible Text-to-Speech API powered by Qwen3-TTS with faster-qwen3-tts CUDA graph acceleration.

7-10x faster than stock inference. Real-time generation on RTX 4060/5060 Ti class GPUs. No flash-attn, no vLLM, no Triton — just CUDA graphs.

Features

• OpenAI-compatible /v1/audio/speech endpoint (JSON body)
• CUDA graph acceleration — 7-10x faster than baseline
• Streaming output ("stream": true returns chunked WAV)
• Voice cloning from 3-second reference audio
• 10 languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian
• 9 built-in voices (CustomVoice model) with instruction-based emotion control
• No flash-attn dependency required
• Supports RTX 50-series (Blackwell) GPUs

Quick Start

docker run -d --gpus all \
  -p 8080:8080 \
  -v /mnt/user/appdata/qwen3-tts-api/models:/root/.cache/huggingface \
  -e MODEL_ID=Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \
  --shm-size=4g \
  --name qwen3-tts-api \
  ghcr.io/hsiang-han/qwen3-tts-api:latest

Or with docker compose:

docker compose -f docker/gpu/docker-compose.yml up -d

First start downloads model (~3-7GB) and captures CUDA graphs on first request. China users: set HF_ENDPOINT=https://hf-mirror.com.

Usage Examples

# Generate speech with built-in voice
curl -X POST http://localhost:8080/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input": "Hello, this is a test.", "voice": "Vivian", "language": "English"}' \
  --output output.wav

# With emotion instruction (1.7B CustomVoice only)
curl -X POST http://localhost:8080/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input": "我真的太开心了！", "voice": "Vivian", "instruct": "用特别开心的语气说"}' \
  --output happy.wav

# Streaming output
curl -X POST http://localhost:8080/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input": "Streaming test.", "voice": "Vivian", "stream": true}' \
  --output stream.wav

# List available voices
curl http://localhost:8080/v1/voices

Built-in Voices (CustomVoice model)

API Endpoints

Environment Variables

Available Models

Hardware Requirements

• NVIDIA GPU with 4GB+ VRAM (0.6B) or 8GB+ VRAM (1.7B)
• NVIDIA driver 550+ (Ampere/Ada) or 570+ (Blackwell RTX 50-series)
• Docker with NVIDIA Container Toolkit

Credits

• Qwen3-TTS by Alibaba Qwen Team — the model
• faster-qwen3-tts by @andimarafioti — CUDA graph acceleration (7-10x speedup)