Falling pitch and slowing rate can help decide that talk has ended

What it is

Utterance ends often carry prosodic change: fundamental frequency (F0) falls or becomes a boundary tone, the last syllables lengthen, rate slows, amplitude drops. These cues can help the “are they done?” decision, especially when silence is still short and the semantics are incomplete. They are suprasegmental acoustics, not the words themselves, and not a raw energy gate.

Why it happens

Talkers mark chunk boundaries for listeners with prosody. Final lowering on statements, rising tunes on questions in some languages, and pre-boundary lengthening are stable tendencies. An energy-only endpointer treats a brief low-energy dip mid-sentence as the end, or a last syllable that still has voicing as unfinished. Adding F0 trajectory, duration, and spectral tilt to the classifier gives the machine the same kind of evidence listeners use while waiting for the next phrase. Language and individual differences are large: some people lower pitch mid-clause; some stay high at the end. A question rise treated as “not done yet” holds the question open. Prosody is therefore probabilistic evidence to be weighted with silence and semantics, not a single veto.

Studying it

Train endpointers with and without prosodic features on speech with marked boundaries; compare time error at true ends and false boundaries in the middle. Material must include questions, lists, and unfinished trail-offs. Replicate across languages, because in tone languages F0 also carries lexical tone and final lowering is more crowded. Task (command versus narrative) and L2 talkers belong among independent variables. Broadcast speech has exaggerated prosodic boundaries compared with conversation.

Where it stops holding

Whisper, narrowband telephony, and very low SNR make F0 tracking fail; the prosodic channel dies first. In tone languages, lexical tone overlaps boundary tone, and F0 alone will treat a tonal minimal pair as a boundary. Synthetic or very flat read speech has little usable prosody. Writing “can help” as “submit on a fall” will cut unfinished lists (“apples, bananas, …”).

Applying it

• Treat a prosody score as a third evidence stream, weighted with silence and semantics, and regress separately on questions and lists.
• Re-estimate F0 features on the target language; do not ship English statement-fall rules onto Chinese.
• When F0 tracking is unreliable (whisper, noise), fall back to silence plus semantics automatically, and mark the prosody channel as failed in logs.

Related

• Same group: C7.09.1 Endpointing can use silence duration alone or jointly with semantic completeness · C7.09.3 A fixed silence threshold behaves inconsistently across speaking rates · 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: boundary tone · pre-boundary lengthening · prosodic endpointing