Final contour does the turn-yielding
Aliases: boundary tone · turn-yielding cue · terminal pitch
What it is
The system says “I found three.” with a falling end; it sounds finished. The user says “ok,” overlapping “the first is….” If the synthesizer closes every sentence on a declarative fall, it opens a transition-relevance place at every period. Turn-yielding final contour is the pitch shape on the last syllable of system speech — fall, rise, or level hold — telling the listener whether they may speak. It is about how the sentence currently being spoken ends, not about filling the two silent seconds before the system starts.
Why it happens
In talk, yielding cues travel as a bundle: a fall or a question rise, a drop in intensity, final lengthening. Listeners project “I may come in” on those cues; they do not wait for dead air. Default TTS boundary tones are almost always falling, so every syntactic period becomes a joinable gap — even when the system still has the next list item or a second slot to speak. Continuing material that ends on a level or slight-rise hold marks the gap as “still mine,” and premature incoming drops. A question that should yield needs a rise to hand over the second pair-part; a fall on “red or blue” sounds like the choice is already made, and people may not answer.
Pause length cannot stand in for contour. The same hundred milliseconds of inter-sentence silence is heard as a hand-over after a fall and as a breath after a hold. Making every ending the same fall, then forbidding barge-in in the state machine, still sounds like “you finished and will not let me talk.”
Studying it
Manipulate boundary tone on otherwise identical wording: fall, rise, level hold. Measure user incoming relative to the end, whether incoming lands where the system still meant to continue, and whether people answer after questions. Dependent measures are premature-onset rate and legal uptake after questions — not “naturalness” scores. A natural-sounding fall can still yield in the wrong place.
Conversation-analyse real TTS: at each syntactic boundary mark “system still holds / has yielded,” against user onset. The densest premature yields sit between list items and in the middle of multi-slot readbacks. A lab that substitutes a button for speaking will not see floor fights.
Where it stops holding
One-way playback (a timer chime, an alarm) never hands over the floor; contour is not a yielding problem there. Experts who already know a prompt’s rhythm will not barge in on a mid-item fall — that fails for new users and rare skills. In noise the pitch shape is inaudible and yielding has to move onto words (“still speaking,” “your turn”). Treating final contour as “make it more human” sends work into timbre and misses its job in turn-taking.
Applying it
- Periods where the system will continue (between list items, mid multi-slot readback) get a level or slight-rise hold, not a completed fall. Falls and rises are for genuine questions and completions.
- Do not default to “falling close, then the next item.” To continue, mark the hold in the melody first.
- A question’s ending must be hearable as a question. A falling choice is a floor not handed over.
- How to check: contrast boundary tones on lists and multi-slot readbacks. Premature incoming should be higher on mid-item falls than on holds; uptake after questions should be higher on rises than on falls.