Contracts Bridge Audio
Transport: ws://127.0.0.1:8778
Envelope: same as Bridge ↔ Kiosk
Status: SPEC
| type | data | meaning |
|---|---|---|
wake |
{score, model} |
Wake word fired above threshold. |
vad.speech_start |
{} |
Speech onset detected. |
vad.speech_end |
{durationMs} |
vad_silence_ms of silence elapsed. |
audio.begin |
{turn_id, sampleRate, encoding} |
Precedes the binary stream. |
audio.chunk |
binary frame | 16 kHz mono PCM s16le, ~320 ms per frame. |
audio.end |
{frames, totalMs} |
Stream complete. |
level |
{rms} |
~10 Hz mic level for the UI meter. |
error |
{code, message} |
Device lost, model load failure. |
| type | data | meaning |
|---|---|---|
listen.start |
{turn_id, maxMs} |
Begin streaming captured audio. |
listen.stop |
{} |
Stop capture (timeout, cancel, barge-in). |
wake.enable |
{enabled} |
Disabled during playback in half-duplex mode. |
play |
{audioUrl, turn_id} |
Unused in Phase 1 — reserved; kiosk is the sink (ADR 0001). |
mute |
{muted} |
Software mute. Reflected in mic_status MCP tool. |
- Rolling in-memory ring buffer (~3 s) — never written to disk by default.
- Include ~500 ms of pre-roll before the wake word in the captured utterance.
- Hard stop at
max_utterance_seconds. - STT is batch after VAD end, not streaming into the model mid-utterance (product path).
Live accuracy path:
mic → comstar-audio (VAD) → bridge PCM → Reach SpeechClient or HttpSttClient → transcript
