Endpointing is when the system decides the user has finished
Aliases: end-of-turn · floor taking after endpoint · utterance-complete decision
What it is
Someone says “remind me to call… uh… mom tomorrow at three.” If the system starts talking after “call,” the words were not the failure — the floor was taken too soon. Endpointing as turn yield is the claim that the decision “this utterance is over” is, in dialogue, the system announcing it now holds the speaking right. How voice activity is sliced on the input path is a separate primitive. What matters here is who stands on the floor after the slice, and how collisions are heard.
Why it happens
Human turn-taking uses transition-relevance places: syntax can close, pitch falls, a yieldable gap appears. Machines approximate that with a silence timeout or a “query complete” score. Once the score crosses, the system launches its own turn — a question, a confirmation, a read-back before acting. Cut early, and the user’s turn-constructional unit is still open: two speakers at once, the user forced to repeat the tail. Cut late, and the user has already yielded; the room holds an unowned blank, like a missed catch.
Endpointing is therefore not a neutral trim of a recording. It is a claim on the floor. The same energy gate shows up in dialogue as interruption or as dead air. The input channel may only report “no speech now.” The dialogue system has to translate “no speech” into “they yielded” or “they are breathing.” Mis-translation is paid at the turn layer: people who were cut off raise volume or switch to shorter, sharper sentences. They do not retune their glottis.
Studying it
Annotate live commands for floor collision: at the instant the system starts audio, is the user still speaking (cutoff), and how large is the gap from a stable user stop to system onset (lag). Align those instants with human-labeled transition-relevance places, not with the energy curve’s own threshold. Dependent measures: cutoff rate, lag in milliseconds, and the repair after cutoff (repeat, barge the system, abandon).
Do not rerun voice-activity ROC on this card. That curve answers “is this speech.” The question here is whether the system’s first sample falls where the other party had already yielded. Under driving or occupied hands, spoken gaps stretch and the collision distribution moves; a “best silence gate” from a quiet desk is not yet a dialogue policy.
Where it stops holding
Streaming recognition can start understanding before endpoint; that changes internal compute, not the fact that when to speak is still a floor claim. User barge-in on system speech runs the other direction and belongs to interruption detection. Two people talking to one device means there is no single yielder. Ultra-short commands (“pause,” “off”) have almost no internal gap; the cost of early versus late endpointing is then skewed toward “a little late is cheaper.”
Applying it
- Treat endpoint timeout as a product-level turn parameter, not a constant buried in the recognizer. For older adults, or slots that need long number strings, default to the system speaking later.
- If the user continues within tens to a little over a hundred milliseconds after system onset, treat it as cutoff: stop your own turn, keep the recognized prefix, wait out the rest. Do not execute the half-utterance as a complete command.
- Accept on collisions, not on word error. Sample sessions for “system audio while user still speaking.” If that rate will not fall, the yield policy is what to change, not the vocabulary.