语音生成服务器

Elise Voice MCP Server

An MCP (Model Context Protocol) server that provides text-to-speech capabilities using the Elise/Ceylia voice dataset from Hugging Face.

Current Status (Pre-MLX Migration)

This commit represents the working state using Coqui TTS XTTS-v2 running on CPU.

Features

• Text-to-Speech Generation: Convert text to speech with automatic playback
• Voice Characteristics: Query detailed information about the Elise/Ceylia voice (pitch, speaking rate, quality metrics)
• Sample Texts: Browse sample texts from the original dataset
• Auto-playback: Generated audio plays automatically via macOS afplay
• Easy Integration: Works seamlessly with Claude Desktop and other MCP clients

Dataset

This server uses the Jinsaryko/Elise dataset from Hugging Face, which contains high-quality speech samples from the speaker "Ceylia" with detailed acoustic annotations.

Installation

Prerequisites

• Python 3.10 or higher
• pip or uv package manager

Setup

1. Clone or navigate to this directory

2. Create a virtual environment (recommended)

   复制

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate
   

3. Install dependencies

   复制

   pip install -r requirements.txt
   

   Or using uv:

   复制

   uv pip install -r requirements.txt
   

Configuration

For Claude Desktop

Add this configuration to your Claude Desktop config file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "elise-voice": {
      "command": "/Users/martin.suehowicz/code/voicemcp_server/venv/bin/python",
      "args": [
        "/Users/martin.suehowicz/code/voicemcp_server/src/elise_voice_server.py"
      ],
      "cwd": "/Users/martin.suehowicz/code/voicemcp_server/src"
    }
  }
}

Note: Update the paths to match your actual installation directory and virtual environment location.

Usage

Once configured in Claude Desktop, you can use the following tools:

1. Generate Speech

Ask Claude to generate speech from text:

Generate speech saying "Hello, I'm Ceylia, and I'm here to help you today!"

This will create an audio file in the audio_output/ directory.

2. Get Voice Information

Query voice characteristics:

What are the characteristics of the Elise voice?

This returns information about pitch, speaking rate, and quality metrics from the dataset.

3. Browse Sample Texts

See what kind of content the voice was trained on:

Show me some sample texts from the Elise dataset

Available Tools

The server provides three MCP tools:

Output

Generated audio files are saved in the audio_output/ directory as WAV files and automatically play via macOS afplay (can be disabled with play: false).

Development

Project Structure

voicemcp_server/
├── src/
│   ├── __init__.py
│   ├── elise_voice_server.py  # Main MCP server
│   ├── voice_dataset.py       # Dataset loader
│   └── tts_engine.py          # TTS generation
├── audio_output/              # Generated audio files
├── requirements.txt
├── pyproject.toml
└── README.md

Running Tests

pytest

Technical Details

• MCP SDK: Uses the official Python MCP SDK for server implementation
• Dataset Loading: Automatically downloads and caches the Elise dataset from Hugging Face
• TTS Engine: Coqui TTS XTTS-v2 (tts_models/multilingual/multi-dataset/xtts_v2)
• Model Location: ~/Library/Application Support/tts/tts_models--multilingual--multi-dataset--xtts_v2
• Audio Format: Generates WAV audio files
• Inference: CPU-only (Apple Silicon MPS support unreliable)
• Performance: ~2-3 seconds per short sentence, ~20 seconds for long sentences
• Dependencies: PyTorch 2.5.1, Transformers 4.39.3 (downgraded for compatibility)

Known Issues

• Voice cloning currently disabled due to torchcodec/ffmpeg compatibility issues
• Apple Silicon GPU (MPS) support unreliable with XTTS-v2
• CPU-only inference is slow for longer text

Next Steps

Planning to migrate to mlx-audio for:

• Better Apple Silicon optimization (Neural Engine + GPU)
• Faster inference
• Voice cloning support via CSM model
• More reliable performance

Troubleshooting

Dataset Loading Issues

If you encounter issues loading the dataset:

• Ensure you have a stable internet connection
• Check that you have sufficient disk space for the dataset cache (~500MB)
• The dataset will be cached locally after first download

Model Download Issues

If XTTS-v2 model fails to download:

• Model is ~6GB, ensure sufficient disk space
• Use manual download: huggingface-cli download coqui/XTTS-v2 --local-dir ~/Library/Application\ Support/tts/tts_models--multilingual--multi-dataset--xtts_v2

Audio Generation Errors

If speech generation fails:

• Verify that TTS library is properly installed
• Check that the output directory is writable
• Ensure XTTS-v2 model is fully downloaded
• Check logs: ~/Library/Logs/Claude/mcp-server-elise-voice.log

Claude Desktop Integration

If the server doesn't appear in Claude Desktop:

• Verify the config file path is correct
• Check that the cwd path in the config matches your installation
• Ensure virtual environment paths are absolute
• Restart Claude Desktop after changing the config
• Check Claude Desktop logs: ~/Library/Logs/Claude/mcp-server-elise-voice.log

License

MIT License

Acknowledgments

• Dataset: Jinsaryko/Elise on Hugging Face
• TTS Engine: Coqui TTS
• MCP Protocol: Anthropic's Model Context Protocol