Local - Piper

Local text-to-speech via Piper, a fast neural TTS system optimized for local execution.

Official Documentation

• Piper GitHub
• Wyoming-Piper GitHub
• Voice Samples
• Voice Models on Hugging Face
• PyPI - piper-tts

---

Local Setup

Using Docker (Recommended)

Wyoming-Piper Server:

docker run -d \
  --name piper \
  -p 10200:10200 \
  -v /path/to/voices:/data \
  rhasspy/wyoming-piper \
  --voice en_US-lessac-medium

With HTTP Server (for REST API):

# Using piper-http-server
pip install piper-http-server
piper-http-server --port 5002 --model en_US-amy-medium.onnx

Manual Installation

1. Download Piper binary:

# Linux x86_64
wget https://github.com/rhasspy/piper/releases/latest/download/piper_linux_x86_64.tar.gz
tar -xzf piper_linux_x86_64.tar.gz
cd piper

2. Download a voice model:

# Download voice model and config
wget https://huggingface.co/rhasspy/piper-voices/resolve/main/en/en_US/amy/medium/en_US-amy-medium.onnx
wget https://huggingface.co/rhasspy/piper-voices/resolve/main/en/en_US/amy/medium/en_US-amy-medium.onnx.json

3. Run HTTP server:

pip install piper-http-server
piper-http-server --port 5002 --model en_US-amy-medium.onnx

Using pip (Python)

pip install piper-tts

# Test from command line
echo "Hello world" | piper --model en_US-amy-medium.onnx --output_file test.wav

Verify

# Test TTS endpoint
curl -X POST http://localhost:5002/synthesize \
  -H "Content-Type: application/json" \
  -d '{"text": "Hello world"}' \
  --output test.wav

# Play the audio
ffplay test.wav  # or: aplay test.wav

---

Provider Configuration

import { PiperTTSProvider } from '@llmrtc/llmrtc-provider-local';

const tts = new PiperTTSProvider({
  baseUrl: process.env.PIPER_URL || 'http://localhost:5002'
});

Environment Variables

Variable	Default	Description
PIPER_URL	http://localhost:5002	Piper server URL

Provider Options

interface PiperConfig {
  baseUrl?: string;   // Server URL
  voice?: string;     // Voice model name
}

---

Available Voices

Voice	Language	Quality	Size
en_US-amy-medium	English (US)	Good	60MB
en_US-lessac-medium	English (US)	Good	60MB
en_US-ryan-medium	English (US)	Good	60MB
en_GB-cori-medium	English (UK)	Good	60MB
de_DE-thorsten-medium	German	Good	60MB
es_ES-mls_9972-medium	Spanish	Good	60MB
fr_FR-siwis-medium	French	Good	60MB

Browse all 30+ languages at rhasspy.github.io/piper-samples.

---

Key Features

• Fast Inference: Optimized for real-time use, even on Raspberry Pi
• ONNX Models: Uses efficient ONNX-based VITS voice models
• Multi-language: 30+ languages supported
• Small Footprint: Voice models are typically 60-100MB
• No Internet Required: Fully offline operation

---

Notes

• Choose a fast Piper voice (medium quality) for real-time use
• Works well with streaming TTS enabled in the server
• Both .onnx model and .json config files are required
• Python package supports 3.7-3.10
• For lowest latency, run on a machine with good single-thread CPU performance