Automatic Analysis and Synthesis of Fujisaki's Intonation Model for TTS

This paper deals with the automatic analysis and synthesis of intonation using Fujisaki’s model. Both the accent commands and the phrase commands are related to the accent group. We propose an analysis method which imposes strong linguistic constraints. This method gives good results when compared to other current methods and is the best option for synthesis (at least when using accent groups). For synthesis, several prediction algorithms are evaluated. The results show that VCART (an extension of CART to predict vector values) gives the best performance when compared with standard CART or with Neural Networks. The paper also analyzes which features are more relevant to predict the parameters of Fujisaki’s model.