The goal of this project is to determine how turn-taking in human-human dialogue works, and use this as a basis for developing a flexible model of turn-taking for use in spoken dialogue systems. The first two aims focus on determining whether turn-taking is solely determined by the current speaker or is negotiated by both conversants. The first aim measures how well human subjects, listening to excerpts of speech, can predict whether the current speaker will continue the turn or release it. The second aim analyzes timing lags at turn transitions. Many turn transitions have short lags, which favors a speaker-control model; however this might be an artifact of common speech act sequences, back channels, early onsets, and provisional turns. The third aim, using the findings of the first two aims, is to determine the cues and mechanisms that are used by conversants in taking the turn.
Better understanding of turn-taking in human-human dialogues is important as it will help in building more efficient and natural spoken dialogue systems, allowing the user and system to better collaborate to solve complex tasks. It should also allow spoken dialogue systems to deal with a broad range of users, from experts to novices: for experts, letting them take the initiative (and the turn), in order to efficiently complete the task, while guiding novices with more directions and examples. Furthermore, understanding human-human turn-taking might have biomedical applications. For example, Autism, which is a disorder that affects social communication, might impact how people engage in turn-taking, and so turn-taking biomarkers might help in diagnosing it.
