Qwen3-TTS MCP Server

A Model Context Protocol (MCP) server that exposes Qwen3-TTS voice synthesis capabilities with Voice Design.

Features

• Advanced Voice Synthesis: Generate realistic audio using the Qwen3-TTS 1.7B model
• Voice Design: Customize voice with natural language descriptions
• Multilingual: Support for 10 languages (Auto, Chinese, English, Japanese, Korean, French, German, Spanish, Portuguese, Russian)
• MCP Integration: Access via Model Context Protocol for integration with LLMs

Installation

pip install -e .

Usage

As MCP Server

Start the MCP server:

python main.py

The server will connect via stdin/stdout and expose the following tools:

Available Tools

1. generate_tts: Generate audio from text

   • text: Text to convert to speech (single string or list for batch)
   • language: Language (Auto, Chinese, English, etc.)
   • voice_description: Description of desired voice characteristics
   • max_tokens: Maximum generation tokens (default: 2048)

   Example (single):

   {
     "text": "Hello, how are you?",
     "language": "English",
     "voice_description": "deep male voice, friendly tone"
   }
   

   Example (batch):

   {
     "text": ["Hello, how are you?", "こんにちは、元気ですか？"],
     "language": ["English", "Japanese"],
     "voice_description": ["deep male voice, friendly tone", "cheerful female voice"]
   }
   

2. generate_tts_voice_clone: Generate audio using voice cloning from reference audio

   • target_text: Text to convert to speech with cloned voice
   • language: Language (Auto, Chinese, English, etc.)
   • ref_audio_base64: Reference audio encoded in base64 (WAV, 3-10 seconds recommended)
   • ref_text: Transcript of reference audio (required if use_xvector_only is False)
   • use_xvector_only: If True, use only speaker embedding (faster but less accurate)
   • model_size: Model size ("0.6B" or "1.7B", default: "1.7B")
   • max_tokens: Maximum generation tokens (default: 2048)

   Example:

   {
     "target_text": "Hello, how are you today?",
     "language": "English",
     "ref_audio_base64": "UklGRi4...",
     "ref_text": "Okay. Yeah. I resent you. I love you.",
     "use_xvector_only": false,
     "model_size": "1.7B"
   }
   

3. list_languages: List all supported languages

MCP Architecture

The server implements the MCP specification with:

• Tools: Available tools for clients to call
• Resources: (Future) Access to files and data
• Prompts: (Future) Reusable prompts for TTS

Requirements

• Python >= 3.12
• CUDA (recommended for better performance)
• GPU with sufficient memory (4GB+)
• Optional: flash-attn for improved performance with flash attention

Project Structure

.
├── main.py              # MCP server implementation
├── pyproject.toml       # Project configuration
├── requirements.txt     # Python dependencies
└── README.md            # This file

Performance Optimizations

This implementation uses optimizations from the official Qwen3-TTS repository:

• Flash Attention 2 for faster inference and lower memory usage
• bfloat16 precision for efficient computation
• Batch inference support for processing multiple requests

Next Steps

• [ ] Add support for MCP resources
• [ ] Implement generated audio caching
• [ ] Add structured logging
• [ ] Create example client
• [ ] Deployment documentation