Whisper + Piper Voice Pipeline

Local STT (speech-to-text) and TTS (text-to-speech) as a single HTTP server. Zero cloud dependencies.

Architecture

Audio In → POST /transcribe → Whisper (faster-whisper) → JSON {text, language}
Text In  → POST /speak       → Piper TTS → ffmpeg → audio/ogg (Opus)

Both endpoints run in one Python process on one port (default: 9998).

Quick Start

1. Install dependencies:

python3 -m venv ~/whisper-env && source ~/whisper-env/bin/activate
pip install faster-whisper
apt install ffmpeg  # or brew install ffmpeg on macOS

1. Download Piper + a voice:

mkdir -p ~/piper && cd ~/piper
wget https://github.com/rhasspy/piper/releases/latest/download/piper_linux_x86_64.tar.gz
tar xzf piper_linux_x86_64.tar.gz
mkdir voices && cd voices
wget https://huggingface.co/rhasspy/piper-voices/resolve/main/de/de_DE/thorsten_emotional/medium/de_DE-thorsten_emotional-medium.onnx
wget https://huggingface.co/rhasspy/piper-voices/resolve/main/de/de_DE/thorsten_emotional/medium/de_DE-thorsten_emotional-medium.onnx.json

1. Run the server (scripts/voice-server.py):

python3 voice-server.py --port 9998 \
  --whisper-model small --whisper-device cpu \
  --piper-bin ~/piper/piper/piper \
  --piper-model ~/piper/voices/de_DE-thorsten_emotional-medium.onnx

API

Transcribe (audio → text):

curl -X POST -F "file=@message.ogg" http://HOST:9998/transcribe
# {"text": "Hallo Welt", "language": "de"}

Speak (text → audio):

curl -X POST -H "Content-Type: application/json" \
  -d '{"text": "Hallo Welt", "speaker": "4"}' \
  http://HOST:9998/speak -o response.ogg

Configuration

Choosing Models

Whisper: small for CPU (good balance), medium for GPU (best quality without large-v3 overhead).

Piper voices: Browse https://rhasspy.github.io/piper-samples/ — download .onnx + .onnx.json files.

Full Setup Guide

Read references/setup-guide.md for systemd service config, all voice options, model comparison table, and OpenClaw integration details.