Tree-Guided Transformation-Based Intonational Phrase Break Prediction

In Text-to-Speech (TTS) systems, the prediction of intonational phrase breaks is important for both the naturalness and intelligibility of synthetic speech. This chapter proposes a hybrid algorithm called the tree-guided transformation-based learning (TGTBL), which combines the classification and regression tree (CART) with the transformation-based learning (TBL), to predict the intonational phrase breaks from the unrestricted text. It automatically generates the TBL templates, thereby minimizing the need for human supervision during the TBL training. Results of comparative experiments show that, for the task of intonational phrase break prediction, templates automatically generated by CART achieve comparable performance to manually summarized templates, and make an improvement of 10.18% on the F-score over the atom templates.