📄 Melodic Cues to Turn-Taking in English: Evidence from Perception

Summary

The paper asks whether listeners use syntax or pitch contour to decide if a Southern British English speaker is about to give up the floor. The authors took two Cambridge Map Task dialogues from the IViE project, about 30 minutes in total, and split them into inter-pausal units (IPUs) using a 100 ms pause threshold. Each boundary was tagged three ways: same speaker or speaker change, syntactically complete or not, and one of five final contours (H* H%, H* %, HL H%, HL L%, H*L %).

Two perception tests followed. In the first, listeners heard a fragment and guessed what would happen next. In the second, they heard the real exchange and judged what the first speaker had intended.

Syntax turned out to be the main cue. Listeners almost never expected a turn change after an incomplete clause, and no contour reliably signalled yielding. The exception was the high level tone H* %, which signalled holding the floor whatever the syntax. This replicates earlier Dutch findings and points to cross-language differences in how backchannels are invited.

Key points

• 25 native Southern British English listeners (9 men, 16 women, aged 19–54) took both tests. Experiment One used 38 randomised stimuli, each played twice. Experiment Two used 20 exchanges, four per contour.
• When syntax was incomplete, only 2% of Experiment One predictions were a turn change. When syntax was complete, H* % still drew 89% "hold" predictions.
• The truncated fall (HL %) favoured continuation much more than the full fall (HL L%), with 71% versus 41% turn-keeping responses.
• In Experiment Two, only 8% of judgements credited the speaker with wanting to yield after H* %. Once a real change was heard, roughly a third of incomplete-syntax fragments ending in a fall or fall-rise were judged as intended yields.
• High rises were rare in the English data. Backchannel expectation after H* % was 6% in English against 56% reported for Dutch.

Interesting ideas

• I1. The stimuli were chosen to cross syntactic completeness with turn outcome and contour, so each cue's separate effect and their interaction could be estimated instead of being confounded.
• I2. The same listeners judged the same material twice: once predicting the outcome blind, and once judging with the outcome known. This exposed a hindsight effect, where seeing the outcome pushed judgements toward a cooperative reading.
• I3. The truncated fall also lacks a high or low boundary tone, yet it did not keep the turn. Using it as a contrast case ruled out the simpler explanation that any missing boundary tone means "hold".
• I4. High-rise results were split by sentence type: interrogatives drew "change" while declaratives drew "backchannel". A label driven by sentence mood can thus masquerade as an effect of another cue.
• I5. Holding the floor and inviting only a backchannel were merged into one "continuation" class, on the grounds that a backchannel does not take the floor.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Evaluation sets could be built as factorial grids, varying one cue such as tense, negation or clause completeness while holding others fixed, so each feature's real contribution can be measured.
• I2 (learn): When LLM teacher tiers label silver data, compare labels given with and without the downstream answer visible, to detect teachers that rationalise the outcome.
• I3 (tree): When a conversion rule or repair rule seems explained by one surface property, look for a minimal contrasting construction that shares that property, to falsify the shortcut before adopting the rule.
• I4 (frames): Speech-act and turn-projection logic should condition on clause mood first, so other cues are not credited with what mood already explains.
• I5 (frames): In dialogue frames, short acknowledgements like "uh-huh" can be modelled as non-turn acts that leave the current speaker's discourse frame open, rather than as new turns.