the voice assistant kept saying no speech detected when there was clearly speech. first suspect was the wake word, so the threshold got lowered and the refractory period shortened, which changed nothing. the actual test was the giveaway: say the wake word, get a response, say something into the follow up listen, and the first listen transcribed in a couple hundred milliseconds while the second one came back completely empty.
the logs nailed it. the speech to text call fired, succeeded, and returned empty at the exact same second the assistant logged a binary message for a non existing handler. translation: on the follow up listen, the phone app streams your audio into a pipeline that already tore down. your voice arrives at a dead websocket door and never reaches the router, so the engine hears nothing at all. that is why the threshold fix did not help — the wake word never fails, the re listen stream dies client side.
the proof the speech chain is fine: pulled the saved debug audio from the day before, real speech with a strong signal, and it transcribed perfectly through the same pipeline. the first utterance of the day went through the identical chain flawlessly. the difference is the trigger — the phone still hands off to the system assistant for the wake word, and the companion app's re listen path fights that same handoff. the stream dies exactly there. the fix is to point the assist trigger at the microphone button for testing, or configure the native wake word instead of the system handoff.