Analysis and Prediction of Myanmar Prosodic Unit Boundary

Myanmar belongs to the Lolo-Burmese sub-branch of the Tibeto-Burmese branch of the Sino-Tibetan language family and is a tonal language. In the front-end text analysis of speech synthesis, the prosodic structure analysis and unit's boundary prediction are crucial to the naturalness of speech synthesis. In order to improve the naturalness of Myanmar speech synthesis, this paper studies prosodic features and prosodic unit boundary prediction. The size of prosodic units and the duration of syllables before and after their boundaries have been studied in this paper. To realize automatic prosodic unit boundaries labeling, a method of labeling based on the combination of word segmentation text and silence duration is proposed. Based on BiLSTM-CRF model, we also have designed and implemented a method to predict the boundaries of prosodic units from Myanmar text. Finally, the boundary prediction results are applied to the speech synthesis system based on HMM to evaluate its naturalness. The experimental results show that our method of automatic prosodic boundary labeling and prosodic unit boundary prediction can improve the naturalness of speech synthesis.