Modeling pitch contour of Chinese Mandarin sentences with the PENTA model

Abstract

In continuous speech, the pitch contour of the same syllable may vary much due to its contextual information. The Parallel Encoding and Target Approximation (PENTA) model is applied here to Mandarin speech synthesis with a method to predict pitch contours for Chinese syllables with different contexts by combining the Classification And Regression Tree (CART) with the PENTA model to improve its prediction accuracy. CART was first used to cluster the syllables' normalized pitch contours according to the syllables contextual information and the distances between pitch contours. The average pitch contour was used to train the PENTA model with the average contour for each cluster. The initial pitch is required with the PENTA model to predict a continuous pitch contour. A Pitch Discontinuity Model (PDM) was used to predict the initial pitches at positions with voiceless consonants and prosodic boundaries. Initial tests on a Chinese four-syllable word corpus containing 2048 words were extended to tests with a continuous speech corpus containing 5445 sentences. The results are satisfactory in terms of the Root Mean Square Error (RMSE) comparing the predicted pitch contour with the original contour. This method can model pitch contours for Mandarin sentences with any text for speech synthesis.