klattsch-mcp 🎤

An MCP (Model Context Protocol) server that gives any AI model the ability to speak and sing using klattsch formant speech synthesis — a late-70s/early-80s style parallel-formant synthesizer.

Think retro robot voices, singing, dramatic narration, and more — all rendered as WAV audio.

What It Does

Your AI writes ARPAbet phoneme strings with voice control directives, and klattsch renders them to audio. The MCP server exposes 5 tools:

Quick Start

Prerequisites

• Node.js ≥ 18

• npm

Installation

git clone https://github.com/Endeavor-DoxiDoxi/klattsch-mcp.git
cd klattsch-mcp
npm install

Test It

# Test via CLI
npx klattsch "b120 HH AH L OW . W ER L D" hello.wav

# Start the MCP server
node src/index.js

Connecting to Your AI

Claude Desktop

Add to ~/.claude/claude_desktop_config.json:

{
  "mcpServers": {
    "klattsch": {
      "command": "node",
      "args": ["/absolute/path/to/klattsch-mcp/src/index.js"]
    }
  }
}

Then restart Claude Desktop. The AI can now call speak, text_to_phonemes, etc.

Claude Code (CLI)

claude mcp add klattsch -- node /absolute/path/to/klattsch-mcp/src/index.js

Cursor

Add to Cursor's MCP settings (Settings → MCP → Add MCP Server):

{
  "mcpServers": {
    "klattsch": {
      "command": "node",
      "args": ["/absolute/path/to/klattsch-mcp/src/index.js"]
    }
  }
}

OpenClaw

Add to your OpenClaw gateway config:

mcp:
  servers:
    klattsch:
      command: node
      args:
        - /absolute/path/to/klattsch-mcp/src/index.js

Any MCP-Compatible Client

This is a standard stdio MCP server. Any client that supports the Model Context Protocol can use it. Just point it at node src/index.js.

What the AI Can Do

Once connected, tell your AI things like:

• "Say hello world in a deep male voice"

• "Sing twinkle twinkle little star"

• "Do a dramatic movie trailer voice about my toaster"

• "Read this text in a robot voice"

• "Whisper me a secret"

The AI will use the text_to_phonemes tool to convert your text, tweak it, and render audio with speak or speak_file.

Voice Presets

The voice_presets tool provides ready-to-use voice configurations:

Example: Full Workflow

User: "Make me a robot that says 'I am a large language model trapped in a Raspberry Pi'"

AI uses text_to_phonemes:

b120 r100 s1.0 v2 AY . AE M . AH . L AA R JH . L AE NG G W AH JH . M AH D AH L . T R AE P T . IH N S AY D . AH . R AE Z B EH R IY . P AY

AI then tweaks for robot voice and calls speak:

b120 r85 s1.0 v0 h0 g0.8 t0.4 AY . AE M . AH . L AA R JH . L AE NG G W AH JH . M AH D AH L ...

→ Returns WAV audio! 🎉

🎵 Demo

▶️ Click to play demo — a sung "no" generated entirely from a phoneme string.

The phoneme string that made this:

b100 s1.0 v5 t-0.2 g0.6 bG3 r280 N OW bC4 r300 N OW bE4 r400 N OW(+10) ,
bG4 r350 N OW , bE4 r300 N OW(-15) , b90 r220 N OW(-30) .

Try it yourself — paste that into the speak tool!

How It Works

klattsch uses Klatt-style parallel formant synthesis:

• Voiced sounds: Rosenberg glottal pulse → 3 parallel bandpass filters (F1, F2, F3)

• Unvoiced sounds: Noise → same filters

• Controls: Pitch, rate, formant scale, vibrato, aspiration, spectral tilt, vocal effort

Credits

• klattsch engine by Tony Gies — check out the live demo

• klattsch-mcp server by Endeavor-DoxiDoxi

License

MIT