Myanmar Prosodic Features Analysis And Prediction For Speech Synthesis

The process of using computers and other equipment to convert text into sounds that people can understand is called speech synthesis.With the rapid development of artificial intelligence applications,it is possible to realize intelligent dialogue with computers.Speech synthesis is the core technology for realizing human-computer voice interaction.Myanmar is the standard language,the language spoken by all ethnic groups and people in various regions in Myanmar.Compared with common languages such as English and Chinese,Myanmar speech synthesis research is still relatively lagging.The naturalness of speech synthesis,especially the expressiveness of prosodic rhythm,needs to be further improved.This thesis will aim to effectively improve the naturalness of Myanmar speech synthesis,study the prosody characteristics of Myanmar,and explore the prediction methods of prosody features.The main work of the thesis includes:（1）Based on the phonetic database and the language characteristics of Myanmar,we analyze the acoustic performance where the Myanmar prosodic units are at the boundaries and is before or after the boundaries using a large number of “text-speech”pairs.The purpose is to explore the relationship between text features and prosody features.（2）A method for automatically marking the boundaries of prosodic units that combines text and speech features is proposed.We use four models of HTK,CRF,Bi LSTM and Bi LSTM-CRF to automatically label the boundaries of Myanmar prosodic units,and apply the labeling method to the speech synthesis database and the training of speech synthesis acoustic models.（3）A method for automatically predicting the boundaries of prosodic units through Myanmar text is proposed.We use four models of CRF,Bi LSTM,Bi LSTM-CRF and BERT-CRF to predict the boundaries of prosodic units,and apply the prediction method to the text be synthesized so that the corresponding synthesized speech has the proper prosodic feature performance.（4）Designed and implemented a Myanmar speech synthesis system incorporating prosodic features.The HMM-based and HMM-DNN-based speech synthesis systems are implemented respectively.Then in the thesis,we also apply the prosody labeling and prosody prediction to the front-end text analysis and processing of the synthesis system to verify the effectiveness of the methods proposed.The experimental results show that the acoustic performance at the boundary of the prosody is a sign of the division of the prosodic unit boundary;the automatic labeling prosodic units boundaries based on CRF model has the best performance,and meets the requirements of corpus construction for Myanmar speech synthesis;the BERT-CRF model has the best performance,and is suitable for the analysis of the front-end text of speech synthesis,which can obviously improve the naturalness of Myanmar speech synthesis.