Why architecture matters here

Wake word fails on false accepts (unintended activations) and false rejects (missed calls). Architecture matters because model + threshold + user tuning shape UX.

Advertisement

The architecture: every piece explained

The top strip is the path. Microphone. VAD. KWS model. Trigger.

The middle row is quality. False accepts. False rejects. User tuning. Privacy.

The lower rows are practice. Multi-word / hotword. Metrics. Ops — updates + language + battery.

Wake word — KWS models + always-on + false accepts + user tuningthe word that starts the conversationMicrophonealways on ring bufferVADvoice presence gateKWS modelsmall NNTriggerscore + thresholdFalse acceptsfrictionFalse rejectsuser says againUser tuningpersonal voicePrivacyon-device onlyMulti-word / hotworde.g. 'Hey X'MetricsFAR + FRROps — model updates + language + batterytriggerrejectadaptprotectcomposemeasuremeasureoperateoperate
Wake word detection: mic → VAD → KWS → trigger.
Advertisement

End-to-end flow

End-to-end: mic ring buffer sampled. VAD active. KWS model runs on windows. Threshold triggers wake. Rest of pipeline activates. Privacy: audio processed on-device only until trigger.