Voice

Provides voice tools for Claude Code with on-device Whisper STT and Supertonic TTS for fast, local voice interactions.

python

0

GitHub Stars

python

Language

6 months ago

First Indexed

3 months ago

Catalog Refreshed

Documentation & install

Readme and setup notes from the catalogue, plus a client-ready config you can copy for your MCP host.

Installation

Add the following to your MCP client configuration file.

Configuration

View docs

{
  "mcpServers": {
    "jochiang-voice-mcp": {
      "command": "uv",
      "args": [
        "run",
        "--directory",
        "/path/to/voice-mcp",
        "voice-mcp"
      ]
    }
  }
}

You can run a local MCP server that adds voice capabilities to Claude Code. It handles on-device speech-to-text with Whisper and on-device text-to-speech with Supertonic, enabling hands-free interactions and faster voice conversations without external API calls.

How to use

To use the Voice MCP Server, connect it to Claude Code as an MCP server. You will have access to a set of voice tools for input and output: listen_and_confirm to record and transcribe your speech for confirmation, listen_for_yes_no for quick binary decisions, speak to synthesize and play back text, and combined tools like speak_and_listen and speak_and_confirm to reduce latency in back-and-forth conversations.

Practical usage patterns include initiating a spoken query with listen_and_confirm, then Claude repeats the transcript for your verification. For quick yes/no questions, use listen_for_yes_no. Request spoken responses with speak, or engage in a full spoken interaction by using speak_and_listen or speak_and_confirm.

How to install

Prerequisites you need before starting are Python 3.10 or newer and the uv package manager. You also need a microphone for speech-to-text and speakers or headphones for text-to-speech.

Step 1: Install the MCP server dependencies and run the server locally using uv.

Step 2: Add the MCP server configuration to Claude Code so it can load the voice tools during startup.

Configuration and usage notes

The Voice MCP Server runs locally and does not require external API calls for transcription or synthesis. Speech-to-text uses Whisper locally, and text-to-speech uses Supertonic locally. You can adjust silence handling and beeps to suit your microphone and environment to improve recording reliability.

Server presence is configured as a standard MCP portal within Claude Code. Default behaviors include auto-downloading model data on first run and on-device audio processing for responsiveness.

Troubleshooting

No audio output from TTS: verify your output device and consider enabling stereo output if your DAC requires it.

Recording stops too quickly: adjust the SILENCE_THRESHOLD and SILENCE_DURATION_S values in the audio configuration to suit your mic. Default values are SILENCE_THRESHOLD = 0.01 and SILENCE_DURATION_S = 2.5 seconds.

Recording doesn’t stop fast enough: reduce SILENCE_DURATION_S to shorten the cutoff window.

Notes

First-run loads Whisper small model and Supertonic on demand. The Whisper model runs on CPU by default, with options to switch to CUDA for GPU acceleration if your system supports it.

Available tools

listen_and_confirm

Record speech, transcribe with Whisper, and return the transcript for your confirmation.

listen_for_yes_no

Listen for a quick yes or no answer and interpret it as yes, no, or unclear.

speak

Synthesize and play back the provided text using the on-device TTS engine.

speak_and_listen

Speak a prompt and then listen for a full response in a single operation to reduce round-trip latency.

speak_and_confirm

Speak a prompt and then listen for a yes/no confirmation in a single operation to reduce round-trip latency.