A probabilistic approach to prosodic word prediction for Mandarin Chinese TTS

Prosodic word is a basic rhythmic unit of Mandarin Chinese Speech. It is one of the most important factors determining the naturalness of the generated speech by a TTS system. This paper investigates the problem of predicting Chinese prosodic words from word sequence. First, we examine the patterns of Chinese prosodic words and investigate the key features for prediction. Then a baseline model of CART is used. Based on this model, the effects of the number of POS categories and the number of single word categories are investigated. Finally, a Markov chain approach is proposed. This model has the advantages of both CART approach and other statistical approaches, while the drawbacks of those approaches are avoided. Experiment shows that the proposed Markov chain approach outperforms the simple CART approach.