Speech Recognition Of Hainan Dialect Based On Deep Learning

Hainan has been a special economic zone for 31 years,and keep digging with its strengths.With three flags of the special economic zone,international tourism island and ecological environment,Hainan got a rapid development of regional social economy.The proportion of gross output value of the tertiary industry increased year by year,with tourism contributes a lot to the total output value of the tertiary industry,the number of tourists continues to grow,tourism revenue has performed well,but based on a small number of constraints,there is still a lot of room for improvement.One of the most important reason is the cultural difference caused by the language barrier,especially because Hainan is an island with its unique dialect,which makes it difficult for people from outside the province to learn,thus restricting the development of islanders in Hainan.Speech recognition technology is an important means to alleviate this problem.This paper is based on a large number of studies on Hainan dialect,choosing Hainan min,the dialect with the largest number of speakers in Hainan,as the research object.The most representative words were selected to establish the corpus of Hainan dialect.Based on the study of dialect pronunciation,a pronunciation dictionary of Hainan dialect is designed,volunteers were then invited to participate in the recording of dialect speech.After segmentation,speech enhancement methods such as adding noise and speed disturbance were used to expand corpus data,the first Hainan dialect speech corpus was built at last.Then based on the current popular speech recognition technology CTC,this paper compares the performance of different networks LSTM with DFSMN in corpus,the best performance network DFSMN is used as the acoustic modeling method of the speech recognition system.A speech recognition system of Hainan dialect based on deep learning is preliminarily established.It verifies the feasibility of deep learning in speech recognition of Hainan dialect.