Wyoming-OpenAI-Gateway

A gateway that exposes local Wyoming protocol services via OpenAI-compatible API endpoints. This allows any OpenAI-compatible client to use local Wyoming speech services (TTS and STT) without modification.

Features

• OpenAI-Compatible API — Drop-in replacement for OpenAI's /v1/audio/speech, /v1/audio/transcriptions, /v1/audio/translations, and /v1/voices endpoints
• Wyoming Protocol — Connects to any Wyoming-compatible TTS service (Piper, Sherpa-ONNX, etc.) and any Wyoming-compatible STT service (faster-whisper, Sherpa-ONNX, etc.)
• Streaming Support — Real-time audio streaming for both TTS and STT protocols
• Zero System Dependencies — Audio transcoding via miniaudio (bundled wheels, ~30x faster than subprocess-based ffmpeg)
• Health Checks — Kubernetes/Docker-ready /healthz and /readyz endpoints

Quick Start

Docker

docker run -d \
  --name wyoming-openai-gateway \
  -p 8555:8555 \
  -e TTS_HOST=your-tts-server \
  -e TTS_PORT=10200 \
  -e STT_HOST=your-stt-server \
  -e STT_PORT=10300 \
  ghcr.io/archmonger/wyoming-openai-gateway:latest

Using pip

pip install wyoming-openai-gateway

TTS_HOST=127.0.0.1 TTS_PORT=10200 \
STT_HOST=127.0.0.1 STT_PORT=10300 \
wyoming-openai-gateway

Using Docker Compose

wget https://raw.githubusercontent.com/archmonger/Wyoming-OpenAI-Gateway/main/compose.yml
TTS_HOST=127.0.0.1 TTS_PORT=10200 \
STT_HOST=127.0.0.1 STT_PORT=10300 \
docker compose up -d

API Reference

GET /v1/voices

Lists available TTS voices from the Wyoming server.

Response:

{
  "voices": [
    {
      "id": "en_US-lessac-medium",
      "name": "Lessac (Medium)",
      "languages": ["en_US"]
    }
  ]
}

POST /v1/audio/speech

Generates speech audio from text.

Request Body:

{
  "input": "Hello, world!",
  "model": "wyoming",
  "voice": "en_US-lessac-medium",
  "response_format": "wav",
  "speed": 1.0,
  "stream": false
}

Response: WAV audio data (Content-Type: audio/wav)

POST /v1/audio/transcriptions

Transcribes audio to text using a Wyoming STT server.

Request: multipart/form-data

Response:

{
  "text": "Your hands lay open in the long fresh grass."
}

POST /v1/audio/translations

Same as transcriptions, but translates the audio to English text.

Configuration

All configuration is done via environment variables:

Validation: At least one of TTS_HOST or STT_HOST must be defined. It is an error to set only a host without its corresponding port, or vice versa. If neither host is defined, the application will exit with an error.

Architecture

┌─────────────────────┐       ┌──────────────────────┐       ┌──────────────────────┐
│                     │       │                      │       │                      │
│  OpenAI-compatible  │ HTTP  │  Wyoming-OpenAI-     │ TCP   │  Wyoming Protocol    │
│  Client             │──────▶│  Gateway             │──────▶│  Server (TTS/STT)   │
│  (curl, ST, Home   │       │                      │       │                      │
│   Assistant, etc.)  │       │  FastAPI → Wyoming   │       │  Piper / Whisper     │
│                     │◀──────│  Protocol Translator  │◀──────│                      │
└─────────────────────┘       └──────────────────────┘       └──────────────────────┘

The gateway acts as a protocol translator between OpenAI's HTTP API and the Wyoming TCP protocol. It maps REST endpoints to Wyoming events (Describe, Synthesize, Transcribe, etc.) and translates audio data formats transparently using miniaudio.

Development

git clone https://github.com/archmonger/Wyoming-OpenAI-Gateway.git
cd Wyoming-OpenAI-Gateway

# Install with dev dependencies
pip install -e ".[test]"

# Run tests
pytest --cov=wyoming_openai_gateway

# Run linting
ruff check src/ tests/

# Start the gateway
TTS_HOST=127.0.0.1 TTS_PORT=10200 \
STT_HOST=127.0.0.1 STT_PORT=10300 \
python -m wyoming_openai_gateway

License

MIT