Joint extraction and prediction of fujisaki's intonation model parameters

This paper presents a joint extraction and prediction framework for intonation modeling applied to Fujisaki’s intonation model for text-to-speech conversion. Previous methods in the area extract the parameters of accent and phrase commands for each sentence. Then, these parameters are related to linguistic features for prediction. In our approach commands that share the same linguistic features are globally estimated. This approach intends to overcome some consistency problems of the extracted model parameters. The global nature of the parameter optimization avoids the interpolation step, which sometimes can produce a bias in the extracted parameters. Experimental results show that the higher consistency of the parameters result in a higher accuracy when the fundamental frequency contours are predicted.