Prompt: Build Your Voice System

Copy-paste this into Claude Code to set up Chatterbox TTS with voice cloning and integrate it with OpenClaw.

Prerequisites

Mac with Apple Silicon (M1/M2/M3/M4) for MPS acceleration (works on Intel too, just slower)
Python 3.11+ and uv installed
OpenClaw running at localhost:18789
A short audio sample (15-60 seconds) of the voice you want to clone
- Must be clean audio, minimal background noise
- WAV or MP3 format

Part 1: Install Chatterbox TTS

# Install uv if you don't have it
curl -LsSf https://astral.sh/uv/install.sh | sh

# Create a directory for TTS
mkdir ~/tts && cd ~/tts

# Create a virtual environment
uv venv
source .venv/bin/activate

# Install Chatterbox
uv pip install chatterbox-tts

# Test it works (no voice clone, just default voice)
python -c "
from chatterbox.tts import ChatterboxTTS
import torchaudio
model = ChatterboxTTS.from_pretrained(device='mps')  # 'cpu' if no Apple Silicon
wav = model.generate('Hello, this is a test.')
torchaudio.save('test.wav', wav, model.sr)
print('Success! Check test.wav')
"

Part 2: Clone a Voice

# Clone a voice from your audio sample
python -c "
from chatterbox.tts import ChatterboxTTS
import torchaudio

model = ChatterboxTTS.from_pretrained(device='mps')

# Path to your audio sample (15-60 seconds)
VOICE_SAMPLE = '/path/to/your/voice-sample.wav'
VOICE_NAME = 'myvoice'

wav = model.generate(
    'This is a test of my cloned voice. Testing one two three.',
    audio_prompt_path=VOICE_SAMPLE,
    exaggeration=0.5,  # 0.0 = subtle clone, 1.0 = strong exaggeration
)
torchaudio.save(f'{VOICE_NAME}-test.wav', wav, model.sr)
print(f'Voice clone test saved to {VOICE_NAME}-test.wav')
"

# Listen to it — adjust exaggeration (0.3-0.7 is usually good range)
open myvoice-test.wav

Part 3: Build the TTS Server

Use this prompt with Claude Code:

I need a Chatterbox TTS server that acts as an ElevenLabs-compatible API,
so OpenClaw can use it for text-to-speech.

## Requirements

Build a Python FastAPI server that:

1. Accepts requests in ElevenLabs format:
   POST /v1/text-to-speech/{voice_id}
   Body: {"text": "...", "model_id": "..."}
   Returns: audio/mpeg (MP3 file)

2. Supports multiple voice clones (one per voice_id)

3. Uses Chatterbox TTS with MPS acceleration (Apple Silicon)

4. Starts fast (model loads once on startup)

5. Handles concurrent requests (async)

## Voice Configuration
Create a voices.json config file:
{
  "voices": {
    "myvoice": {
      "sample_path": "/path/to/sample.wav",
      "exaggeration": 0.5,
      "display_name": "My Voice"
    },
    "default": {
      "sample_path": null,
      "exaggeration": 0.3,
      "display_name": "Default"
    }
  }
}

## Tech Stack
- Python with FastAPI
- uv for package management
- Chatterbox TTS (already installed in ~/tts/.venv)
- Run on port 4126
- Start with: python server.py

## File Structure
~/tts/
├── server.py          # FastAPI server
├── voices.json        # Voice configurations  
├── samples/           # Voice sample WAV files
│   └── myvoice.wav
├── cache/             # Generated audio cache (optional)
└── .venv/

## Extra Features
- Health check endpoint: GET /health
- List voices: GET /v1/voices
- Basic request caching by text+voice_id (optional, saves API calls)
- Logging with request duration

## After building
I'll test with:
curl -X POST http://localhost:4126/v1/text-to-speech/myvoice \
  -H "Content-Type: application/json" \
  -d '{"text": "Hello from my cloned voice"}' \
  --output test.mp3 && open test.mp3

Part 4: Connect to OpenClaw

After the server is running, configure OpenClaw:

// In ~/.openclaw/config.json, add:
{
  "tts": {
    "provider": "elevenlabs",
    "baseUrl": "http://localhost:4126",
    "apiKey": "local",
    "defaultVoice": "myvoice"
  }
}

Then per-agent voice configuration:

# In your agent YAML:
id: chief
voice: myvoice   # matches the voice_id in your voices.json

Test it:

openclaw tts test "Hello, this is my AI assistant speaking."

Part 5: Auto-Start on Login

Help me set up a macOS LaunchAgent to auto-start my Chatterbox TTS server
when I log in.

The server runs as: cd ~/tts && .venv/bin/python server.py
Port: 4126
Log file: ~/tts/chatterbox.log

Create:
1. ~/Library/LaunchAgents/com.myai.chatterbox.plist (LaunchAgent config)
2. ~/tts/start.sh (start script with proper env)

The LaunchAgent should:
- Start on login
- Restart if it crashes
- Log stdout and stderr to ~/tts/chatterbox.log

Load it:

launchctl load ~/Library/LaunchAgents/com.myai.chatterbox.plist

Voice Quality Tips

Sample quality > sample length. A clean 20s sample beats a noisy 2-minute one.
Exaggeration 0.3-0.5 works for most voices. Higher = more distinctive but potentially robotic.
Multiple samples for the same voice (different sentences, same speaker) improves consistency.
Public domain voices: Many historical speeches are in the public domain and make excellent samples.

Ethical Considerations

Only clone voices you have permission to use
Never use cloned voices to impersonate real people in deceptive contexts
Cloned celebrity voices for personal productivity (your own ears only) is generally acceptable
For public presentations: use original samples or truly synthesized voices