Hands-free
Leave the microphone open and answer without touching anything.
Push to talk is the default because it is the honest one: it hears nothing until you hold the key. Hands-free is for the times your hands are somewhere else, and it is the mode where the hard part is not starting a turn on everything else in the room.
- 1
Turn on hands-free in the menu bar
The microphone stays open from then on. Push to talk still works exactly as it did; this is in addition to it, not instead.
- 2
Say the wake word, then talk
A small detector listens for it and nothing else is transcribed until it fires. That is the whole reason it is a wake word and not just an open microphone.
- 3
Stop talking
A run of silence ends the turn and the words go where you were pointing, the same as a held key would have sent them.
- 4
Turn it off when you need the room back
The same Hands-free item in the menu bar, clicked again. The microphone is closed, not ignored, and the notch says so rather than leaving you to trust it.
Off means off
Turning hands-free off closes the device rather than throwing away what it hears. There is nothing listening and nothing buffered, which is the only version of an off switch worth having on an app whose whole argument is that your voice never leaves the room. Push to talk still works while it is off.
Muting is a different switch and works the other way round. The Muted submenu silences one project’s spoken summaries while it keeps working, and it still asks you questions, because an agent blocked on an answer has nowhere else to put it. Muting says nothing about the microphone.
When it goes wrong
It starts on the person next to me
The wake detector is tuned for a quiet desk, not an office.
Raise EARSHOT_WAKE_THRESHOLD above 0.7. It trades a little responsiveness for not transcribing somebody else's conversation. Push to talk is the honest answer in a genuinely busy room: it hears nothing at all until you hold the key.
It never hears the wake word
The threshold is too high for your voice or your microphone.
Lower EARSHOT_WAKE_THRESHOLD a little at a time. Below about 0.5 it starts opening on words that merely rhyme.
A turn ends while I am still thinking
700 ms of quiet ends a turn, and a pause mid-sentence is quiet.
Raise EARSHOT_VAD_SILENCE. Two seconds suits somebody who thinks out loud.