Abstract
Speech dialogue systems, such as Apple's “Siri,” have gradually become more widespread, and in the near future, a greater number of general users will have the opportunity to communicate with such systems. To facilitate this, it is necessary for the system to communicate more naturally with users, and to realize that both verbal and nonverbal information must be taken into consideration. In this research, using an available speech recognition platform, we have developed simple dialogue system that consider utterance timing and evaluated it. As a result, it is clarified that the mean error of generated pause duration is controlled within 20msec, and the proper fixed pause duration is turned to be 600msec in this system. Moreover, the evaluation of generative models of pause duration showed that the model that synchronized the utterance duration weakly or fixed pause duration (600msec) model tended to get best results.
