Implementation And Improvement Of Malay Speech Synthesis System

With the rapid development of technologies such as speech synthesis and speech recognition,human-machine voice interactive applications are becoming more and more popular.How to make the machine emit the sound consistent with the real person has always been the goal pursued by speech technology researchers.Deep learning based on DNN(deep neural network)has become an effective way to improve the performance of speech synthesis.So far,the results of speech synthesis research that has been achieved have mainly focused on languages such as English and Chinese,and the research results of other languages need to be further enriched.Malay language belongs to the Malay-Polynesian language family of the South Island language and is widely used in Malaysia,Singapore,Brunei and other places.In order to develop the Malay speech synthesis application system,this paper uses HMM and DNN to explore ways to improve the performance of speech synthesis system.The main work of the thesis is:(1)According to the writing and pronunciation features of Malay language,the Malay pronouns are automatically segmented by the structure of phonetic pros and the phoneme structure,and the phoneme structure is selected as the modeling of Malay speech synthesis based on the results of automatic segmentation.(2)According to the linguistic features of Malay,the full-context lab are designed,and the automatic generation of the context attribute set is realized.Then the Malay speech synthesis system based on HMM and full-context lab is designed and debugged,which realizes the training of the model and the synthesis of speech.(3)Aiming at the problem of the decision tree acoustic model based on HMM speech synthesis system,the fully connected deep neural network is used instead of the decision tree as the acoustic model,and the system training and speech synthesis are re-executed.Compared with the HMM speech synthesis system,the synthesized speech quality is more obvious.(4)For the speech synthesis system based on DNN acoustic model,the parameter optimization generation standard of training phase and synthetic parameter generation stage is inconsistent and the fundamental frequency trajectory over-smoothing problem of synthesized speech is adopted.The trajectory training considering global variance is adopted,and the synthesized speech is effectively solved.Smooth the problem and get closer to the natural soundtrack in terms of detailThe experimental results show that the speech synthesis technology based on DNN acoustic model and its global variance trajectory training can further improve the HMM speech synthesis system and effectively improve the synthesized speech quality.