Engagement after a wake must expire
Aliases: listening window · follow-up · session timeout · re-arm
What it is
After a wake-word hit, the device enters a time-bounded engagement interval: the microphone keeps handing audio to recognition, the user can issue a command immediately, and sometimes a few seconds of follow-up are allowed without repeating the phrase. That interval must expire. When it does, the system returns to keyword spotting until the next explicit entry. A listening window that never ends turns a one-shot passphrase into standing authorization.
Why it happens
Engagement exists because a dialogue turn does not finish in the frame where the wake word lands: the user still has to assemble the next sentence, and the system may ask for a slot. Engineering boxes that grant with a listening window or session timeout: audio inside the window is treated as a command; audio outside it is only keyword-spotted. Timeout can be a fixed silence, a wall-clock cap, or “a short tail after this dialogue turn ends.” Too short, and the second sentence never arrives; too long, and bystander talk, television, and private muttering are ingested as input. Wake-word-free follow-up relaxes the window; it does not remove the guard. Timeout is not endpointing: endpointing decides whether one utterance is finished and should be committed; timeout decides whether the whole engagement is still valid.
Studying it
Use multi-turn tasks: after the first wake, insert a second sentence at varying delays, insert bystander speech, and insert television speech. Independent variables include timeout duration, whether follow-up is allowed, and whether the task requires a system question. Dependent measures include follow-up success, how often non-user speech inside the window is treated as a command, and whether people still know they are being heard. A single wake–command–done trial hides the timeout design. Gaze or spoken reports can check whether users still think they can keep talking after the interface has returned to idle.
Where it stops holding
Dictation of long text should not be a command window that dies in a few seconds; that is a separately started capture session. In accessibility settings, people who articulate or plan speech slowly need a longer window; a fixed timeout kicks them back to idle. In a car, a steering-wheel button that listens while held and stops on release replaces timeout with press duration. Always-on intercoms and meeting capture do not run a wake-word state machine, so “must expire” does not apply in that form.
Applying it
- Put a hard timeout on engagement, and turn the listening indicator off when it fires, rather than stopping recognition while still looking as if the device is listening.
- If follow-up is allowed, make the remaining window visible as a countdown or an explicit “you can keep talking”; after timeout, require a new wake.
- Accept with bystander speech and delayed follow-ups: the second sentence inside the window must connect; an unwakened sentence outside it must not run.