Intonation modelling for the synthesis of structured documents

This paper describes experiments concerning the prediction of a good intonation for the synthesis of structured documents. The paper extends our previous research in four important aspects: (i) models are trained and evaluated on read text material (no isolated sentences), (ii) the intonation model is evaluated while fully integrated in the entire prosody model chain, (iii) the feature selection process is completely automated, and (iv) the importance of typical text-level features such as text type, text structure and typesetting are investigated. Clearly, human readings of running texts exhibit a much richer intonation than the intonation observed in read isolated sentences. We try to capture this richness in an intonation model that can be learned automatically using data-driven techniques. Our intonation models are RNNs (Recurrent Neural Networks) which are trained from prosodically labelled databases. Objective tests have demonstrated that acceptable intonation models can be constructed in this way, and that text type and text structure are important features whereas type-setting is not.