Hierarchical pitch target model for Mandarin speech

In this study, a hierarchical pitch target model is proposed to analyze the underlying factors of tones and intonation in Mandarin pitch, which can be applied in speech synthesis systems. This model assumes that the surface pitch contour is produced by approximating the sequential pitch targets assigned to the syllables in an utterance, and each pitch target possesses a hierarchical structure. Moreover, the pitch target associated with each syllable can be generated by adding up the tonal pattern of the syllable and the stepped intonation biases contributed by higher level prosodic units, including prosodic words, prosodic phrases, intonation phrases and utterances. The parameters of these factors can be learned from a speech corpus by a gradient method. Finally, the learned tonal patterns of syllables and the intonation patterns from different prosodic layers are illustrated, and the synthesis experiments show the effectiveness of the presented model.