Palabra × Wyoming (Home Assistant / Rhasspy)

A Wyoming protocol server exposing Palabra Realtime STT and TTS as a speech provider for Home Assistant voice assistants (and any other Wyoming client, e.g. Rhasspy 3).

• ASR — transcribe + audio-start / audio-chunk / audio-stop. Audio is forwarded to Palabra as it arrives, at the rate Home Assistant records at (no resampling; other sample widths and channel counts are converted to 16-bit mono). audio-stop sends finalize; the final transcript is returned as transcript.
• TTS — synthesize → audio-start, audio-chunk… (pcm_s16le 24 kHz mono, streamed as generated), audio-stop. Home Assistant's streaming synthesis (synthesize-start / synthesize-chunk / synthesize-stop) is supported: completed sentences are synthesized while the LLM is still writing.
• Discovery — describe returns the ASR model (13 languages) and the voices.

Run

pip install wyoming-palabra
export PALABRA_API_KEY=plbr_...
wyoming-palabra --uri tcp://0.0.0.0:10300

docker run -e PALABRA_API_KEY=plbr_... -p 10300:10300 wyoming-palabra

Options: --language (TTS language when the request does not say; default en), --default-voice (default_low), --voice <id> (repeatable — advertise additional voices, e.g. ones cloned through the Palabra Voices API), --log-level. Environment equivalents: WYOMING_URI, PALABRA_LANGUAGE, PALABRA_VOICE_ID, LOG_LEVEL, plus PALABRA_REGION.

Home Assistant

Settings → Devices & services → Add integration → Wyoming Protocol → host and port of the server (10300). Palabra then appears as a Speech-to-text and Text-to-speech engine under Settings → Voice assistants.

Voices are advertised per language as <lang>-<voice> (en-default_low, de-default_high, …) because a Wyoming synthesize request carries either a voice name or a language, never both. Home Assistant shows only the voices matching the assistant's language; the server derives the synthesis language from the chosen voice, falling back to the language of the last transcribe, then --language.

Tests

pip install -e ".[dev]"
pytest -m "not live"                          # offline: a real Wyoming client against the server with a fake Palabra
PALABRA_API_KEY=plbr_... pytest -m live -s    # describe -> synthesize -> transcribe round trip against the API

Status

Checked against wyoming 1.10 (required Info / Artifact fields, optional Synthesize.voice, one handler per connection) and exercised end-to-end with wyoming.client.AsyncTcpClient — the class Home Assistant uses — against the Palabra API. Not yet run inside a live Home Assistant. Each connection opens its own Palabra STT session with the same API Key, so voice commands from several satellites are transcribed in parallel.