Modeling Pitch Contour of Chinese Mandarin Sentence with PENTA Model

In continuous speech, it is believed that the pitch contour of the same syllable may vary a lot due to its different context information. To apply the Parallel Encoding and Target Approximation (PENTA) model to Mandarin speech synthesis and improve its prediction accuracy, this paper proposed a method to predict pitch contours for Chinese syllables with different contexts by combining the Classification And Regression Tree (CART) with the PENTA model. We first used CART to cluster syllables’ normalized pitch contours according to the context information of syllables and the distances between pitch contours. For each cluster, we calculated the average pitch contour and trained the PENTA model with this average contour. The initial pitch value is required while using PENTA model to predict a continuous pitch contour. We further proposed a Pitch Discontinuity Model (PDM) to predict such initial pitch values at positions where voiceless consonants and prosodic boundaries are found. We first conducted experiments on a Chinese four-syllable word corpus containing 2048 items and then extended experiments to a continuous speech corpus containing 5445 sentences. The results were satisfactory in terms of the Root Mean Square Error (RMSE) values comparing the predicted pitch contour with the original contour. With this method, we can model pitch contours for Mandarin sentences of any text and apply the trained model parameters into speech synthesis.