Intonation Modelling for the Synthes

Human readings of structured documents exhibit a much richer intonation than that observed in read isolated sentences. It is a challenge to capture this richness in an automatic way using datadriven techniques. In this paper, we extend our previous research on intonation modelling for isolated sentences in different respects: (i) the RNN (Recurrent Neural Network) intonation model is now trained and evaluated on read documents, (ii) the model is evaluated as part of the overall prosody model, (iii) the feature selection process is completely automated, and (iv) the importance of textlevel features such as text type, text structure and type-setting are investigated. It is demonstrated that acceptable intonation models can be constructed starting from a database that does not contain any explicit hand labelling of the intonation contours. It also appears that text type and text structure are important features whereas type-setting is not.