Hands-free

Talk to your house without finding a phone

A wake word matched by a model running in this tab, an interruption that actually stops the agent mid-sentence, and a spoken confirmation narrow enough that somebody agreeing with somebody else in the same room can never unlock your front door.

The same thing, as a pipeline

The wake word runs on openWakeWord and the endpointing on silero, both downloaded from three.ws and executed in this tab. That audio is never uploaded. Only the sentence after the wake word is sent, and only to be transcribed.

mic
 -> silero VAD          on device
 -> openWakeWord        on device
 -> /api/asr            the utterance only
 -> agent turn
 -> /api/tts/speak
 -> the agent speaks

The latency budget

Leg	Budget
Wake word detection	200 ms
End of speech	400 ms
Transcription round trip	900 ms
Agent turn	1200 ms
Action to device change	700 ms
First audible reply	1800 ms

Measured live in the panel while you use it, not asserted here.

Why a door is different

Turning a light off happens the moment you ask. Unlocking a door does not: the agent says the whole sentence, the entity appears on screen, and only the word confirm goes through. A general yes does not, and on a device with no screen a guarded action is refused outright.

Every state

The same panel component, driven into each of the twelve states it can be in. Off is the default, and nothing about listening is downloaded until you turn it on.