A study of caller response intervals in spoken dialog systems

One common challenge in spoken dialog systems is to determine how long a system should wait until taking back the turn if the caller is silent. This paper presents the results of an experiment that analyzed the impact of different pause durations on turn-taking issues in a spoken dialog system. The results show that for pauses up to one second long, callers are unlikely to assume that it is their turn to speak, independent of the question type. The results also show that the cumulative response time of the caller is dependent on the type of question they are asked. Additionally, we present the frequency of turn-taking issues as a function of pause duration, as well as recommendations of what pause durations to use in order to minimize turn-taking issues for the two different question types that were analyzed.