A fixed silence threshold behaves inconsistently across speaking rates

What it is

The same silence threshold in milliseconds is not the same event for a fast talker and a slow one. Fast talkers have short inter-word gaps, so a fixed threshold feels sluggish. Slow talkers, older speakers, and L2 speakers often pause longer than the threshold and have one sentence cut in two. The inconsistency comes from gating relative rate with absolute time.

Why it happens

Pause scale tracks syllable rate. At a fast rate, a mid-sentence gap may be only one or two hundred milliseconds; at a slow rate, searching for the next word can pause half a second and still continue. The threshold is wall-clock time and does not read a rate estimate. A 700 ms value tuned on medium lab speech therefore wastes wait for fast users and seals too early for slow ones. Adaptive methods scale the threshold by a recent-window syllable or phone rate, or by a Bayesian update from that user’s observed history. This is related to the planning-pause conflict and not the same axis: planning is content; rate is articulatory rhythm. Someone can be fast and unplanned, or slow and still planning.

Studying it

Have fast, medium, and slow talkers (or the same person instructed to change rate) say the same sentences, sweep a fixed threshold for premature splits and extra wait, then compare a rate-adaptive version. Report stratified by rate, not averaged away. Older and L2 groups should be separate rows. Do not set the threshold only on newscaster rate.

Where it stops holding

Extremely fast speech glues phones together; recognition fails before endpointing gets a turn. Chant, litany, and deliberately lengthened emphatic syllables fool rate estimates. Push-to-talk does not depend on this threshold. Joint semantic endpointing can partly compensate a slow talker’s long gaps, but if recognition also inserts silence frames because of the slow rate, compensation is limited.

Applying it

• Do not serve everyone with one millisecond number; estimate rate per user or per session and scale trailing silence.
• Give slow-speech and accessibility profiles a longer default, and do not hide that notch behind an obscure switch.
• Accept premature splits stratified by rate; a pass on the medium-rate mean is not a pass.

Related

• Same group: C7.09.1 Endpointing can use silence duration alone or jointly with semantic completeness · C7.09.2 Falling pitch and slowing rate can help decide that talk has ended · C7.09.4 Endpoint errors cannot be fixed after the fact, only by respeaking or a manual end
• Adjacent: C7.02 Endpoint Detection · C7.04 Accent, Dialect, and Code-Switching
• Search: speaking rate · adaptive endpointing · silence threshold