atlas/home-voice-agent/tts

TTS (Text-to-Speech) Service

Text-to-speech service using Piper for low-latency speech synthesis.

Features

• HTTP endpoint for text-to-speech synthesis
• Low-latency processing (< 500ms)
• Multiple voice support
• WAV audio output
• Streaming support (for long text)

Installation

Install Piper

# Download Piper binary
# See: https://github.com/rhasspy/piper

# Download voices
# See: https://huggingface.co/rhasspy/piper-voices

# Place piper binary in tts/piper/
# Place voices in tts/piper/voices/

Install Python Dependencies

pip install -r requirements.txt

Usage

Standalone Service

# Run as HTTP server
python3 -m tts.server

# Or use uvicorn directly
uvicorn tts.server:app --host 0.0.0.0 --port 8003

Python API

from tts.service import TTSService

service = TTSService(
    voice="en_US-lessac-medium",
    sample_rate=22050
)

# Synthesize text
audio_data = service.synthesize("Hello, this is a test.")
with open("output.wav", "wb") as f:
    f.write(audio_data)

API Endpoints

HTTP

• GET /health - Health check
• POST /synthesize - Synthesize speech from text
  • Body: {"text": "Hello", "voice": "en_US-lessac-medium", "format": "wav"}
• GET /synthesize?text=Hello&voice=en_US-lessac-medium&format=wav - Synthesize (GET)
• GET /voices - Get available voices

Configuration

• Voice: en_US-lessac-medium (default)
• Sample Rate: 22050 Hz
• Format: WAV (default), RAW
• Latency: < 500ms for short text

Integration

The TTS service is called by:

1. LLM response handler
2. Conversation manager
3. Direct HTTP requests

Output is:

1. Played through speakers
2. Streamed to clients
3. Saved to file (optional)

Testing

# Test health
curl http://localhost:8003/health

# Test synthesis
curl -X POST http://localhost:8003/synthesize \
  -H "Content-Type: application/json" \
  -d '{"text": "Hello, this is a test.", "format": "wav"}' \
  --output output.wav

# Test GET endpoint
curl "http://localhost:8003/synthesize?text=Hello" --output output.wav

Notes

• Requires Piper binary and voice files
• First run may be slower (model loading)
• Supports multiple languages (with appropriate voices)
• Low resource usage (CPU-only, no GPU required)

Voice Selection

For the "family agent" persona:

• Recommended: en_US-lessac-medium (warm, friendly, clear)
• Alternative: Other English voices from Piper voice collection

Future Enhancements

• Streaming synthesis for long text
• Voice cloning
• Emotion/prosody control
• Multiple language support