Design And Improvement Of Burmese Speech Synthesis System

With the development of information technology,speech synthesis technology is more and more widely used,and the research and application of speech synthesis in general languages such as Chinese and English have become popular.Burmese is the official language of Myanmar,belonging to the Tibetan-Burmese family of the Sino-Tibetan language family.Due to its lack of electronic language resources,and speech synthesis research is lagging behind.This paper aims to develop a speech synthesis application system in Burmese,study Burmese text analysis methods,speech waveform synthesis methods,and explore new methods to improve the naturalness of synthetic speech.The main work of the paper is:(1)Normalization of non-standard words in Burmese.The non-standard words such as the special forms of numbers commonly used in Burmese texts,special symbols in the text,and some common abbreviated characters are summarized,and corresponding normalization schemes are given and implemented.(2)Study the pronunciation of Burmese,especially the pronunciation change,design the grapheme-to-phoneme conversion schemes,and design the corresponding rules for the stacked words and complex pronunciation change problems in Burmese.Based on this,a phrase-based statistical machine translation method is introduced to further improve the accuracy of grapheme-to-phoneme conversion.(3)Based on the language characteristics of Burmese,the initials and finals are selected as the Burmese speech synthesis unit(phoneme),and the automatic annotation method of prosodic text including word boundary,syllable boundary and other information is designed.On this basis,the automatic segmentation of phonemes is realized.(4)Based on the language features and syllable formation rules of Burmese language,the context attribute and question set are designed.Based on this,a HMM-based Burmese speech synthesis baseline system is built to implement acoustic model training and speech synthesis.(5)To solve the problems existing in the HMM speech synthesis baseline system,the DNN acoustic model is introduced to replace the decision tree model in the HMM baseline system to improve the accuracy of the acoustic model,and then,aiming at the problems of over smooth fundamental frequency and inconsistent parameter criteria in the speech synthesis system,the global variance trajectory training based on DNN is introduced to further improve the naturalness of speech synthesis.The experimental results show that the Burmese text processing method and HMM-based speech synthesis system designed and implemented in this paper are feasible.Based on this,two DNN-based methods are reasonably introduced to further improve the performance of the baseline system and effectively improve the naturalness of speech synthesis.