Planning When to Say: Dissociating Cue Use in Utterance Initiation Using Cross-Validation

In conversation, turns follow each other with minimal gaps. To achieve this, speakers must launch their utterances shortly before the predicted end of the partner's turn. We examined the relative importance of cues to partner utterance content and partner utterance length for launching coordinated speech. In three experiments, Dutch adult participants had to produce prepared utterances (e.g., vier, "four") immediately after a recording of a confederate's utterance (zeven, "seven"). To assess the role of corepresenting content versus attending to speech cues in launching coordinated utterances, we varied whether the participant could see the stimulus being named by the confederate, the confederate prompt's length, and whether within a block of trials, the confederate prompt's length was predictable. We measured how these factors affected the gap between turns and the participants' allocation of visual attention while preparing to speak. Using a machine-learning technique, model selection by k-fold cross-validation, we found that gaps were most strongly predicted by cues from the confederate speech signal, though some benefit was also conferred by seeing the confederate's stimulus. This shows that, at least in a simple laboratory task, speakers rely more on cues in the partner's speech than corepresentation of their utterance content.