Toward improving dialogue coordination in spoken dialogue systems

Spoken Dialogue Systems (SDSs) often lack natural coordination mechanisms used in human conversation, leading to inefficient interactions. This research investigates verbal and non-verbal cues that support dialogue coordination and explores their application to human-computer interaction. Studies examined differences in speech directed to computers versus humans, finding louder speech for computer-directed utterances, and identified gaze as a key—but error-prone—cue for addressee perception. Analysis of disfluencies revealed that “um” is listener-oriented, while “uh” is speaker-oriented. Further, inter-turn gap length influenced response timing and disfluency likelihood. Finally, reinforcement learning simulations demonstrated that incorporating coordination cues can improve SDS dialogue policies.