On the prediction of global F0 shape for Japanese text-to-speech

The global F₀ shape of Japanese speech is predicted by phrasal accent attributes and adjacent phrasal environment using three-layered neural nets. Three F₀ values of each minor phrase are used for the global shape description, and their prediction is carried out in each major phrase determined by right-branching syntactic boundaries. Through prediction experiments using short and ordinary sentence samples, it is quantitatively confirmed that the global F₀ shapes are predicted fairly well in both samples and that additional controls are necessary for finer prediction in the ordinary sentence samples