Prediction of F0 parameter of contextualized utterances in dialogue

In order to synthesize natural spoken dialogue, it is necessary to incorporate dialogue information into generation of the surface sentence and the prosody. This paper describes the prediction of F0 maximum for minor phrases in dialogue based on a two-step predictive method. Special attentions are directed to speci c phrases containing the person's name or the day of the week in the schedule arrangement task in order to narrow the diversity of characteristics of F0 parameters in dialogue. Seven features were identi ed as dialogue information which are useful to predict the F0 parameter. Two D-rule sets derived from the person's name or the day of the week are very similar to one another. They reduce the total prediction errors by about 50% for the data which have much in uence of dialogue context.