Speech Emotion Recognition Of Datong Dialect Based On Deep Learning

China has a long history and splendid culture.In the course of 5,000 years of history,each region has formed its own unique dialect,which are the treasure of culture in our country.As one of the seven major dialect areas in China,the Datong dialect is the northern dialect.The study of Datong dialect is of great significance to the study of northern dialects.This article focuses on the phonetic emotion of Datong dialect.Due to objective reasons such as late start and immature technology,the research results of Datong dialect speech emotion recognition are deficient.Especially in the corpus resources,it is relatively blank.However,there are many potential demands in society and high potential value in research.Research on the Datong dialect is of great significance for the protection of Chinese intangible cultural heritage.Based on the research of Datong dialect,it has a positive effect on the research of Shanxi dialect,even the national dialect and even the world dialect.The main contents of this paper are as follows.(1)Constructed a corpus of phonetic discrete emotions for the Datong dialect.The work mainly includes the collection,collation and revision of text corpora;the collection,pre-processing(noise reduction,annotation)and collation of speech.Manually expanded and preprocessed these data,and finally obtained 12,000 available voice sample data.(2)Contrast with the emotional features extracted by traditional methods,using traditional machine learning methods to compare the accuracy of different feature sets in emotion recognition of the Datong dialect emotional corpus.The results illustrate the effectiveness of the IS09,IS10,and ComParE feature sets in the Datong dialect,and make a reference for the subsequent experimental feature selection.(3)The global and temporal features extracted under the IS10,ComParE feature set were compared and the two types of features were screened using Convolutional Neural Network(CNN)and Long-Short-Time Memory Network(LSTM),respectively.The global and temporal features of Datong's sentiment corpus are fused,and a correlation neural network(CorrNet)based feature fusion method is used to reduce the similarity of global and temporal features and improve the accuracy of speech sentiment recognition for Datong dialect.Although the related research of the Datong dialect speech emotion recognition technology is still in the preliminary stage,and there are many shortcomings and inadequacies in the experiment,the exploration of Datong dialect speech emotion recognition in this paper can provide some experience for research in this field and will play a role in promoting it in the future.