A thinking pause is misread as end of turn

What it is

“Book a table Saturday evening… (glance at a calendar) … for seven.” Those two seconds are not a turn yield. They are a planning pause: the word is still being found, the complement after “evening” is still on the way. If the silence gate is shorter than that search, the system hears a yield and inserts “which restaurant.” Thinking-pause cutoff is an inside-the-turn gap labeled as finished speech. It is not the other person never answering after a question.

Why it happens

Speakers park unfilled pauses, or filled ones (“uh,” “that…”), before rare words, names, numbers, choice points. Those seams can outlast inter-word gaps and many endpoint timeouts. Location is the tell: the pause sits inside a turn-constructional unit, syntax still open — a preposition, a conjunction, an unfinished list. A pause at a transition-relevance place is a yield; a mid-clause pause is thinking while holding the floor. Duration-only endpointing treats the two seams as one.

The dialogue consequence is concrete: the system’s question is often the slot the second half of the utterance was about to give. The user must first handle the stolen turn (stop and listen, or shove back in) and then supply the missing tail. The repair creates new seams, easy to cut again. Filled pauses are more likely to be treated as speech, so an “uh” can keep the floor; the silent planner is the one most often misread.

Studying it

Elicit turns that force lexical search: uncommon restaurant names, eleven-digit numbers, an address that must be read off paper. Code each pause as mid-clause or clause-final, and whether the system spoke into it. The primary measure is false end-of-turn rate on mid-clause pauses; a secondary is whether the next system question asks for a slot already underway in the unfinished turn. Compare silence-duration endpointing with a syntactic-completeness rule; the gap should show up mid-clause, not at clause end — clause end is where a cut belongs.

Older, slower speakers stretch mid-clause seams further; stratify them rather than folding them into a young-adult mean. Do not use read speech: reading barely produces planning pauses.

Where it stops holding

A mid-clause blank that runs to several seconds may be abandonment; waiting further becomes talking into an empty room. Some languages routinely run long mid-turn seams; thresholds do not travel from English conversation. When the user is scanning a screen, the spoken pause rides with gaze; a device that can see focus should not cut on ear alone. Tiny yes/no answers have almost no mid-clause seam; the mechanism does not apply.

Applying it

• When the last token is a preposition, conjunction, number prefix, or an open list, lengthen the yield gate, or require semantic completeness before speaking.
• After a false cut, do not launch a new question. Say “go on,” or listen briefly, so the unfinished turn can resume.
• Sample transcripts at system onset: is the user’s last word a function word. A high rate means thinking is being cut, not a yield being taken.
• For slots that require long numbers or names, default to wider mid-turn tolerance. Do not reuse a yes/no silence gate.

Related

• Same group: M1.08.1 Endpointing is when the system decides the user has finished · M1.08.3 Dead air before the system speaks needs filler
• Nearby: C7.09 Endpoint Detection and End-of-Utterance · M4.09 Recognition Differences for Children and Older Adults · M1.07 Silence and Timeouts
• Search terms: thinking-pause cutoff · planning pause · false end-of-turn