Speaker

Local TTS for AI coding agents — speak responses aloud

Adds voice output to any AI coding agent (Claude Code, Kiro CLI, Gemini CLI, OpenCode, Crush, Amp). Uses kokoro-onnx (82M params) for fast, natural-sounding speech with ~1.5s latency.

All agents integrate via an MCP server that exposes a speak tool. The Kokoro model loads once and stays warm in memory, eliminating cold-start latency on subsequent calls.

Why?

• Auditory channel helps with focus and processing (especially for neurodivergent users)
• Hearing responses spoken reduces "wall of text" overwhelm
• Natural voice quality — not robotic, won't put you off

Quick Start

git clone <your-repo>/speaker.git
cd speaker
./scripts/install.sh

This installs the speak-mcp MCP server and configures any detected AI tools.

Usage

In any agent session:

Platform	Enable	Disable
Claude Code	/speak-start	/speak-stop
Kiro CLI	@speak-start	@speak-stop
Gemini CLI	@speak-start	@speak-stop
OpenCode	@speak-start	@speak-stop
Amp	@speak-start	@speak-stop

Voice is off by default. When enabled, the agent calls the speak MCP tool with its full response (excluding code blocks).

MCP Server

The speak-mcp entry point runs a FastMCP server exposing three tools: speak, list_voices, and speaker_status. The Kokoro model stays warm in memory — first call loads the model (~2s), subsequent calls have ~200ms overhead.

Tool Schema

Field	Value
Name	speak
Parameters	text: str, voice: str = "am_michael", speed: float = 1.0
Returns	str — confirmation or error message

Adding to Any Agent

Add to your agent's MCP config:

{
  "mcpServers": {
    "speaker": {
      "command": "speak-mcp",
      "args": []
    }
  }
}

Then add to the agent's prompt:

The user can toggle voice with @speak-start and @speak-stop.
When enabled, call the speak tool with your full response text.
Exclude code blocks from spoken text.

See docs/agent-install.md for platform-specific configs.

Voices

kokoro-onnx voices follow the pattern {accent}{gender}_{name}:

Voice	Description
am_michael	American male (default) — clear, natural
af_heart	American female — warm tone
af_bella	American female — bright
am_adam	American male — deeper
bf_emma	British female

Speed

Pass speed as a tool parameter (0.5 = slow, 2.0 = fast, default 1.0).

How It Works

1. Agent calls the speak MCP tool with response text
2. MCP server (speak-mcp) synthesizes audio via kokoro-onnx
3. Audio resampled 24kHz->48kHz and played via sounddevice
4. Model stays warm in memory for low-latency subsequent calls

Architecture

Agent (Claude/Kiro/Gemini/...)
  |
  | MCP protocol (stdio)
  v
speak-mcp (FastMCP server)
  |
  | SpeakerEngine (in-process)
  v
kokoro-onnx -> sounddevice -> audio out

Requirements

• Python 3.10+
• uv for installation
• macOS or Linux (kokoro-onnx runs on CPU via ONNX Runtime)

License

MIT

Generated Artefacts

🔍 Explore this project — AI-generated overviews via Google NotebookLM

🎧 Listen to the Audio Overview	Two AI hosts discuss the project — great for commutes
📊 Browse the Slide Deck	Presentation-ready project overview

Generated by notebooklm-repo-artefacts