Prosodic focus control in reply speech generation for a spoken dialogue system of information retrieval

A spoken dialogue system of information retrieval on academic documents has been developed with a special attention to reply speech generation. In order to realize speech reply with its prosodic features properly controlled to express dialogue focuses, a scheme was developed to directly generating speech reply from reply content. When developing the system firstly, a priority was placed on the automatic processing, and prosodic focus was controlled by rather simple rules (original rules). Based on the listening test for the reply speech generated using original rules, new rules were then developed. Through the further listening test, the rules were revised and called the revised rules. The validity of the revised rules was verified through an evaluation experiment. It was also indicated that there existed users' preferences on the intonation of the reply speech.