Speech Recognition Of Hmong Language In Southeast Guizhou

As the rapid development of artificial intelligence,there are a lot of languages can be recognized by computer.Man-machine Chinese communication has been realized in China,and speech recognition technology has been gradually applied to minority languages.Hmong language is one of the main languages of the ancient people in Southwest China.There are some problems in speech research,such as the lack of language and characters,imperfect speech corpus and regional differences.The construction of Hmong speech corpus and speech recognition can effectively inherit and protect Hmong language and culture,and realize the communication between Hmong language and other languages.In this thesis,Hmong speech in Southeast Guizhou Province is taken as the research object.The construction of Hmong speech corpus,the research on isolated Hmong language word recognition and continuous speech recognition of Hmong language are carried out.The main work is as follows:（1）Aiming at the research of Hmong phonetics,such as the lack of language,incomplete phonetic corpus and regional differences.Using Chinese pinyin and characters to label the voice of the Hmong,based on the central dialect phonetic words collected in Hmong minority language LABS of Guizhou university,a systematic speech corpus of the Hmong is constructed,the corpus content contains the common musical,voice recording personnel from southeast of Guizhou Province can speak native accent of the Hmong people around,The corpus covers most of Hmong language phenomena.（2）Aiming at the problem of speech recognition of isolated words in Hmong language.Using Chinese spell to mark the Hmong language speech,a Hmong language isolated word speech recognition model based on Convolutional Neural Network is constructed.The experimental results show that the recognition accuracy of the Hmong language isolated word recognition model is 97% in the same region and 94% in different regions.After merging the Hmong speech data sets in the same region and different regions,the recognition accuracy is 95%.The model performs well in the speech recognition of combined regions,it proves that the model is effective for the isolated Hmong language word speech recognition marked with Chinese phonetic alphabet.（3）Aiming at the problem of continuous speech recognition in Hmong language.A continuous speech recognition model of Hmong language based on transformer is proposed.The model is based on the modeling method of using Chinese characters as recognition labels.It can directly recognize Hmong language speech into Chinese characters without pronunciation dictionary training,which reduces the dependence on linguistic knowledge.The experimental results show that the model can recognize Hmong Language continuous speech with Chinese as recognition labels,and the word error rate of the model is 31%,It solves the problem of continuous speech recognition of Hmong language without words.