Older adults' rate and pauses trigger early turn decisions

What it is

Many older speakers talk more slowly, and leave longer silent gaps mid-clause while they search for a word — not a turn yield, a lexical retrieval pause. An endpointer calibrated on a younger adult pause distribution will mark those gaps as done and take the floor. Once the person is cut off, the second half — the drug name, the address, the number — never reaches recognition. It looks like “older adults have worse ASR.” What happened first is an early turn decision. The mechanism belongs to this population’s timing, not to the thinking pause anyone can hit mid-turn, and not to the generic engineering fact that one fixed silence threshold is imperfect for both fast and slow talkers.

Why it happens

Aging changes time, not only timbre. Articulatory gestures slow, syllables lengthen; silent gaps while searching for a proper name, a drug, a number lengthen too, often inside an unclosed constituent — after a preposition, in a list, after a measure word. The endpointer’s silence gate, energy gate, and sometimes a “sounds like an ending” prosodic classifier, all take their statistics from a calibration set of younger, faster speakers with shorter pauses. The right tail of the older distribution walks straight across that gate. Once the system speaks, the older adult must first handle a stolen turn: stop and listen, or barge back. Barging back makes new gaps, which get cut again. Filled pauses (“um,” “that…”) sometimes hold the floor because energy is still there; silent searchers are the ones treated as having yielded. The cut-off second half is often the rarest, most slot-critical word, so later word-error totals bill an endpointing error to acoustics.

Studying it

Stratify false ends by age: when the system opens, is the syntax complete, is the last word a function word or a broken list. Histogram mid-clause silence by age band and see where the product threshold sits on which band. Independent variables: whether filled pauses are allowed, whether semantic incompleteness delays end-of-turn. Do not stop at word error: count unfinished turns that were cut, then whether the slot asked in the next turn was already in the uncompleted utterance.

Elicitation should use words older adults actually stop to find — drug names, street names, grandchildren’s nicknames — not news read aloud. Read speech has almost no retrieval gaps and will wash the age effect out. Do not fold older samples into a young mean and then report one endpointing accuracy.

Where it stops holding

Rate does not fall monotonically with age: some older speakers are fast, some younger speakers search slowly. The mechanism targets people whose pause distribution sits outside the right tail of the calibration set; age is a common correlate, not a necessary and sufficient cause. Parkinsonian and other dysarthrias stretch gaps further and need clinically oriented endpointing, not only an “older mode.” Very short yes/no answers have almost no mid-clause gap. Older adults who have learned to say “I’m done” as an explicit end can tolerate a wider gate — that is learned compensation, not a distribution that has returned to young-adult shape.

Applying it

• Default a wider mid-clause silence tolerance for older users, and keep listening after prepositions, measure words, and unclosed lists. Do not reuse the young-adult calibrated gate.
• After a false cut, do not open a new question. Use “please continue” or a short listen-again so the rest of the drug name can land.
• Skills that must carry long proper names (medication, addresses) get their own wider end line. Do not share it with “lights on.”
• How to check: from older-adult sessions, mark whether the user’s last word at system onset was a function word or a broken list. If that rate is high, the system is cutting retrieval, not hearing “unclear speech.”

Related

• Same group: M4.09.1 Children's acoustics and syntax both leave adult norms · M4.09.3 A training-data gap cannot be patched with prompts
• Nearby: M1.08 Turns and Floor Management · C7.09 Endpoint Detection and End-of-Utterance · M4.04 Recognition Differences Across Populations
• Search terms: lexical retrieval pause · older-adult endpointing · age-related speaking rate