Using Past Speaker Behavior to Better Predict Turn Transitions

Using Past Speaker Behavior to Better Predict Turn Transitions Tomer M. Sagie Meshorer Master of Science Center for Spoken Language Understanding within the Oregon Health & Science University School of Medicine June 2017 Thesis Advisor: Prof. Peter A. Heeman Conversations are at the core of everyday social interactions. The interactions between conversants are preformed within the realm of a sophisticated and self-managed turn taking system. In human conversations, the turn taking system supports minimal speaker overlap during turn transitions and minimum gaps between turns. Spoken dialogue systems are a new form of conversational user interface that permits users to use their voice to interact with the computer. As such, the turn taking capabilities of SDS should evolve from a simple timeout to a more human-like model. Recent advances in turn taking systems for SDS use different local features of the last few utterances to predict turn transition. This thesis explores using a summary of past speaker behavior to better predict turn transitions. We believe that the summary features represent an evolving model of the other conversant. For example, speakers who typically use long turns will be likely to use long turns in the future. In addition, speakers with more control of the conversation floor will be less likely to yield the turn. As the conversational image of the speaker evolves as the conversation progresses, other speakers might adjust their turn taking behavior in response. We computed two types of summary features that represent the current speaker’s past turntaking behavior: relative turn length and relative floor control. Relative turn length measures the current turn length so far (in seconds and words) relative to the speaker’s average turn length.