wyoming-mlx 0.1.0

Apple-Silicon-native TTS and STT for Home Assistant and OpenAI-compatible clients

wyoming-mlx

Apple-Silicon-native TTS (Kokoro) and STT (distil-whisper) for Home Assistant and OpenAI-compatible clients.

Why?

If you have a Mac on your network, it can be your voice server. wyoming-mlx turns it into a fast, fully local speech-to-text and text-to-speech service:

• Private by construction. Audio never leaves your network — no cloud speech APIs, no per-request pricing, nothing to subscribe to. Models run entirely on your machine.
• Uses hardware you already own. Apple Silicon's GPU and unified memory run Whisper and Kokoro comfortably alongside whatever else the Mac is doing — no dedicated GPU server, no idle power draw of a CUDA box.
• One service, two ecosystems. Home Assistant talks to it natively over the Wyoming protocol (drop-in replacement for wyoming-faster-whisper/wyoming-piper satellites), while anything that speaks the OpenAI audio API — scripts, editors, chat UIs — can use the same instance via /v1/audio/transcriptions and /v1/audio/speech.
• Set-and-forget. Install with Homebrew, run as a launchd service via brew services, and models are fetched once into the Hugging Face cache.

Speech-to-text runs on MLX via mlx-whisper; text-to-speech runs Kokoro on Metal via PyTorch. The real backends therefore require an Apple Silicon Mac. Everything else (config, HTTP API, Wyoming protocol handling, fake backends, tests) is portable, and CI runs on Linux against the fake backends.

Install (Homebrew)

brew tap rnorth/tap
brew install wyoming-mlx

Run it in the foreground with wyoming-mlx, or as a launchd service:

brew services start wyoming-mlx

Logs go to $(brew --prefix)/var/log/wyoming-mlx.log. Apple Silicon only.

Quick start (dev)

mise install
uv sync
uv run pytest

Works on Linux too: CPU-only torch is selected automatically (the fake backends need no GPU). Integration tests against real models are skipped by default; run them with uv run pytest --integration (Apple Silicon only).

Run locally (fake backends)

uv run python scripts/dev_run.py

The dev server uses the API key dev.

Run locally (real MLX backends)

uv run wyoming-mlx

By default it loads:

• distil-whisper-large-v3 (MLX) on Wyoming port 10300 / HTTP /v1/audio/transcriptions
• Kokoro-82M (MLX) on Wyoming port 10200 / HTTP /v1/audio/speech
• HTTP on port 10400 with API-key auth

Models download on first use to the Hugging Face cache.

API keys

HTTP endpoints require a bearer token. Keys are read at startup from ~/.config/wyoming-mlx/apikeys (override with --http-api-keys-file), one key per line, # comments allowed. The file should be mode 0600. If the file is missing or empty, all HTTP requests are rejected with 401.

Note that the HTTP API listens on all interfaces by default (set WYOMING_MLX_HTTP__HOST=127.0.0.1 to restrict it), and GET /v1/models is unauthenticated, matching OpenAI API behaviour.

mkdir -p ~/.config/wyoming-mlx
(umask 077; openssl rand -hex 32 > ~/.config/wyoming-mlx/apikeys)

HTTP API

List models

curl http://localhost:10400/v1/models

Transcribe an audio file

curl http://localhost:10400/v1/audio/transcriptions \
  -H "Authorization: Bearer $API_KEY" \
  -F file=@some.wav

Synthesize speech

curl http://localhost:10400/v1/audio/speech \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input":"Hello there.","voice":"af_heart"}' \
  --output /tmp/out.wav

Home Assistant integration

Settings → Integrations → Wyoming Protocol → Add:

• STT: <host>:10300
• TTS: <host>:10200

No keys, no TLS (HA convention, trusted LAN).

Configuration

Pass --config /path/to/config.toml or set env vars with the WYOMING_MLX_ prefix and __ for nesting (e.g. WYOMING_MLX_HTTP__PORT=10401). See src/wyoming_mlx/config.py for the full schema.

License

Apache-2.0