Turn-taking timing prediction based on incremental estimation of utterance expectation for spoken dialog systems

Abstract

Turn-taking timing prediction method for spoken dialogue systems is proposed and its evaluation is reported. User's utterance to a system is divided into several speech segments by various reasons. A system with simple turn-taking strategy which takes its turn at every user's speech break may cause troubles such as interrupting user's utterance. Recent studies propose predicition methods of end of user's turn to solve this problem. However, precise timing that system should take its turn is hardly discussed. Thus, we extend our former approach for incremental prediction of end of user's turn so that it can predict precise system turn-taking timing. The idea is based on first order delay response to user's unspeech likelihood and apply thresholding to it. The proposed method is evaluated with actual human dialogue corpus.