Endpointing can use silence duration alone or jointly with semantic completeness

What it is

End-of-speech can be decided from acoustics alone: enough silence duration and the turn is sealed. It can also fold in whether the partial recognition already looks like a complete query (semantic completeness, end-of-query). That is a joint decision: only when the sentence is executable in grammar and slots may a shorter silence commit; if it still looks unfinished, wait longer. This is an operational policy choice, not “whether to endpoint.”

Why it happens

A silence-only detector cannot see words. A joint model uses streaming ASR partials to estimate whether this utterance is already an executable query—verb and object present in an imperative, interrogative structure present in a question—and the posterior rises. Acoustic silence remains evidence, but its weight moves with the semantic posterior: 200 ms may suffice on a complete query; an incomplete “look up…” must wait longer. Language-model bias and domain grammars strongly affect that posterior, treating fillers as unfinished or ellipsis as done. The cost is dependence on recognition quality: a wrong hypothesis fools semantic completeness. Silence is not fooled by words, and will still cut on a planning pause.

Studying it

Run silence-only and a joint model on the same spontaneous queries; compare premature commits, late waits, and whether the hypothesis was executable at commit. Hold out sentences with unfilled slots. Domain (short commands versus dictation) is an independent variable. Joint models should also report wrong seals caused by recognition error. Fully read complete sentences as the only corpus make semantic completeness almost always true.

Where it stops holding

Dictated paragraphs are often intentionally incomplete (more is coming); semantic completeness will rush the seal, so dictation should lean acoustic or use an explicit done control. On a tiny fixed command grammar, joint decision adds little. When recognition is very poor, joint is less stable than silence. Semantic completeness is not the speaker’s sense of being done: they may still add a negation.

Applying it

• Use joint decision to speed short commands; turn off semantic pushing for dictation and open speech, keeping silence or an explicit complete key.
• When partials show an obviously unfinished slot (a verb with no object), forbid commit on short silence.
• In A/B tests, watch premature execution and mean wait together; do not optimise only one.

Related

• Same group: C7.09.2 Falling pitch and slowing rate can help decide that talk has ended · 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 · C7.10 Live Partial Results
• Search: end-of-query · semantic endpointing · joint VAD