Research On Speech Synthesis Of Dungan Language

Speech synthesis is an important part of human-computer interaction,with the maturity of speech synthesis.It has been able to synthesis high natural Chinese Mandarin speech.With the development of machine learning,speech synthesis has received more and more attention and successfully applied to many fields.In recent years,more and more attention has been paid to the synthesis of dialects of national languages in addition to Mandarin and English.It effectively improves the diversity of speech interactions that research on the speech synthesis of national languages and dialects.China has a vast territory and a large number of ethnic groups.In addition to the use of Mandarin,people in different regions still have rich dialects and national languages.Dungan language is the national language used by the Dungan people in the Central Asian countries.Dungan language belongs to the Chinese Central Plains Mandarin,which is a special cross-border variant of Chinese.Aimed at above situation,this thesis selects Dungan language Shaanxi branch to study the characteristics,realizing the speech synthesis of Dungan language.The main work and innovation of the thesis are as follows:Firstly,the thesis builds a corpus of Dungan language and analyzes the phonological features and the phonetic features of Dungan language.The corpus contains 2000 sentences,and it covers all initials,finals,tones,common vocabulary and sentence patterns of the Dungan language.The phonetic corpus is recorded by Dungan students who are native speakers of Dungan.Secondly,the thesis designs a Speech Assessment Methods Phonetic Alphabet for Dungan(SAMPA-DG).And the thesis defines a context-sensitive annotation format suitable for Dungan language,which is used to implement the annotation of the context information of Dungan language.Thirdly,the thesis proposes a Hidden Markov Model(HMM)based Dungan language speech synthesis scheme and a deep neural network(DNN)based Dungan speech synthesis scheme.From the experimental results,when the number of corpus of Dungan language training is small,the effect of HMM-based Dungan language synthesis is better.With the increase of training corpus,the DNN-based Dungan language speech synthesis effect is gradually improved and better than HMM-based.