Speaking-rate adaptation for task-based spoken dialogue systems

Spoken dialog systems are used in diverse applications and they help users accomplish various tasks. Contemporary spoken dialogue systems use prerecorded or synthesized prompts which are invariant in speed throughout the application. Such unvarying speaking-rates may not be suitable to different users pursuing varied tasks.; This thesis examines the possibility of predicting the operator speaking-rate at the utterance level in a corpus of billing-support dialogs. It presents some key factors which affect speaking-rate, including factors based on user-state, dialog-state, user-utterance and operator-utterance characteristics.; Based on this analysis it presents a predictive model to govern the operator-side speaking-rate to use in a system for automatically handling these types of dialogs. Factors that were found to be most useful for predicting the operator-side speaking-rate were the type of the current task, the semantic type of operator's utterance and the duration of the operator's utterance.