Pauses carry punctuation and structure

What it is

Speech has no whitespace. Listeners use pauses to recover punctuation and structure in synthetic speech: a comma-sized gap splits coordination; a period-sized gap splits two speech acts. A ticket kiosk that says “insert card select language” with no boundary is heard as one act, “insert-card-select-language,” not insert then choose a language. The pause here is blank space for syntactic parse. It is not a way of chunking digits or addresses into memorable groups, and it is not a floor-yield at the end of the sentence.

Why it happens

Listeners rebuild syntax from silent (or filled) gaps, pre-boundary lengthening, and F0 reset. A gap of about 150–300 ms is often taken as a clause or list-item boundary; a longer gap is taken as a new speech act. TTS that concatenates SSML without breaks, or a neural model that under-pauses, produces a run-on the parser treats as one constituent. A long silence inside a noun phrase, conversely, splits a word. Pause is punctuation in time: too little, and two predicates glue into one job; in the wrong place, and one job splits into two unrelated sentences.

Studying it

Treat SSML break duration as the independent variable. Ask how many jobs the sentence contains and which way a modifier attaches. Materials: attachment ambiguity (“the old men and women sat,” pause after “men” or not) and dual-act prompts (inform then instruct). Dependent measures: structure choice, cloze hits at the clause boundary — not MOS, and not whether the listener seizes the floor.

No transcript while listening. If the commas are on paper, you did not measure auditory punctuation. In the field, stop at a ticket pillar the instant the prompt ends and ask “how many things did it tell you to do.”

Where it stops holding

Dense sports commentary is still segmented by experts — familiarity substitutes for pause. Verse and brand slogans have their own rhythm and should not be authored as menu prompts. Some languages mark structure more with particles than with silence. Dense codes (card numbers, house numbers) have their own grouping habits; do not mix those with this syntactic punctuation. Lengthening every pause makes people think the system has stalled — a timeout problem, not punctuation.

Applying it

• Author explicit breaks at commas and sentence ends. Listen at 1× without the script and check whether two acts still glue.
• If one prompt holds two speech acts (inform, then instruct), separate them with a period-class pause.
• Do not drop a long silence inside a noun phrase to sound “measured.”
• How to check: no transcript; “how many things did it just tell you to do.” Two acts heard as one means the punctuation pause is missing.

Related

• Same group: M3.02.1 Speech rate must be adjustable · M3.02.3 Stress position changes the meaning of a sentence
• Nearby: M3.07 Rate, pauses and prosody · M2.01 Information density of prompts · M1.08 Turns and floor management
• Search terms: pause as punctuation and structure · prosodic boundary · SSML break