speech-mcp-server

Local Stdio Model Context Protocol (MCP) server wrapping the Kokoro ONNX Neural TTS Engine with real-time Voice Vector Algebra, Central Wisconsin Vernacular Lexicon Overrides, and thread-clamped zero-VRAM CPU execution.

"Neural voice vectors mixed mathematically. Zero VRAM overhead. Pure local performance."

---

🔊 Key Architectural Features

• Voice Vector Algebra Engine: Mix multiple voice profiles using linear vector algebra directly in Python: $$\text{Voice} = (\text{am_adam} \times 0.65) + (\text{bm_lewis} \times 0.30) + (\text{am_michael} \times 0.05)$$
• Central Wisconsin Vernacular Dict: Built-in phonetic lexicon mapper enforcing regional dialect pacing and word pronunciations (bag $\to$ bayg, roof $\to$ ruff, creek $\to$ crick, know'm'sayin).
• Zero GPU VRAM Overhead: Operates entirely on CPU via ONNX Runtime Execution Provider. Keeps GPU completely idle (0% load, 0 MB VRAM).
• Thread-Clamped CPU Execution: Throttled to 2 worker threads (OMP_NUM_THREADS = 2) to eliminate CPU spikes while maintaining sub-200ms latency.
• Kokoro ONNX Engine: High-fidelity 24kHz neural TTS with multiple base voice profiles (am_adam, af_bella, bm_lewis, am_michael).
• MCP Stdio Transport: Full Model Context Protocol compatibility for AI agents and LLM tool integration (/speak).

---

🚀 Quickstart

Prerequisites

• Node.js v18+
• Python 3.10+ with kokoro-onnx, onnxruntime, soundfile, numpy

Installation & Build

git clone https://github.com/yavru421/speech-mcp-server.git
cd speech-mcp-server
npm install
npm run build

Usage

python synth.py "Lexicon overrides active for DuckDB and WASAPI." "am_adam*0.65 + bm_lewis*0.30" "0.94"

Configuration (MCP Client)

Add to your mcpServers configuration:

{
  "speech-mcp-server": {
    "command": "node",
    "args": ["<path_to_repo>/build/index.js"]
  }
}

---

🔒 License

MIT