Testing strategies for bridging time-to-content in spoken dialogue systems

Abstract

What should dialogue systems do while looking for information or planning their next utterance? We conducted a study in which participants listened to (constructed) conversations between a user and an information system. In one condition, the system remained silent while preparing a reply, whereas in the other, it “bought time” conversationally, using strategies from previously recorded human interactions. Participants perceived the second system as better at responding within an appropriate amount of time. Additionally, we varied between mid- and high-quality voices, and found that the high-quality voice time-buying system was also seen as more willing to help, better at understanding and more human-like than the silent system. We speculate that participants may have perceived this voice as a better match for the more human-like behavior of the second system.