The power of special characters in prosodicword prediction for Chinese TTS

Prosodic word (PW) prediction in Chinese Text-To-Speech (TTS) can be formulated as a classification problem that one predicts the tag of every character boundary in a sentence is the PW boundary or not. In this paper, a set of new features called special characters are introduced and put into classifiers to address the PW prediction problem. Some characters often appear at the beginning or at the end of a PW, which make them a strong clue of a PWboundary. Besides, quite a lot of PWs have only one character, which make such characters special. We select a set of special single characters, special starting characters, and special ending characters to help predict PW boundaries. Some special lexical words are often taken as PWs, and we collect a list of such words for PW boundary prediction. Decision tree, Supporting Vector Machine (SVM), MultiLayer Perceptron, and Random Forests are employed as the classifiers. Other features like part-of-speech (POS) of characters, word length, etc. are also used for PW prediction. In our experiments, we got 90.5% and 91.3% accuracies on two corpora containing 8, 000 and 1, 349 sentences respectively, which proved the efficiency of the method.