Voice satellite

espos_voice is a Wyoming-protocol satellite: a TCP server (default port 10700) that lets a signalk-wyoming or Home Assistant orchestrator play TTS to the device and stream microphone audio back, with on-device wake-word detection via esp-sr WakeNet.

espOS supplies the protocol, the server and the wake engine. It does not supply an audio codec: the application implements espos_audio::AudioDriver for whatever the board has and passes it in.

#include "espos_audio/audio_driver.h"
#include "espos_voice/wyoming_satellite.h"

static MyBoardAudio audio;          // : public espos_audio::AudioDriver
static espos_voice::WyomingSatellite sat(&audio, cfg);
audio.init();
sat.start();

espos_audio

A header-only contract, deliberately minimal: init(), ready(), sample_rate(), play_pcm(), plus optional capture and mic-probe hooks. espos_audio::NullAudio is a no-op implementation for boards with no audio, so callers never need a null check.

play_pcm() is non-blocking by contract — it copies and enqueues — so it is safe to call from a UI or render task. A driver that blocks there will stall the display.

Targets

esp-sr supports only esp32s3 and esp32p4. On every other target espos_voice registers as an empty component rather than failing the build, so a project can depend on it unconditionally and simply have no voice where the silicon cannot run WakeNet.

Wake word

The satellite owns a WakeEngine directly, so on-device wake is not optional today. Detection pauses the engine for the duration of an utterance and re-arms it afterwards; see the note in wake_engine.cpp about restoring the shared I2S capture clock after playback, which is what stops WakeNet going deaf after the first TTS reply.