Research Of Speech Emotion Recognition Based On Deep Neural Network

Speech emotion recognition is an important part of human-machine interaction.Human voice contains not only words,but also emotion.At present,speech emotion recognition is the key of AI research,and emotion monitoring on people has very important practical significance.At present,in the study of speech emotion recognition,appropriate database is difficult to get,and model structure,emotion feature,emotion recognition algorithm need to do more research to get satisfying result.This paper has made research of feature extraction,feature learning,feature classification algorithm on speech emotion recognition,and has used deep neural network and multi-level classification algorithm for speech emotion recognition,the main works of the paper as follows:1.The paper has built an emotion database,and extracted the features of speech data.Speech emotion preprocess includes endpoint detection,framing,window adding,and pre accentuation.This paper extracts the prosodic features,voice quality features and spectral features of emotion data,including energy,zero crossing rate,MFCC,fundamental frequency,harmonic to noise ratio,and extract the 12 statistical characteristics of these features,including the maximum,mean,linear slope and etc..A total of 384 dimension statistical features have been extracted,and then this paper has contrasted the differences in the classification ability on different emotion features.2.The paper has proposed an improved algorithm for feature learning based on deep neural network,names DBN-DNN Feature(DDF).The effectiveness of the improved feature learning algorithm is proved by SVM.In the paper,4 kinds of commonly used speech emotion feature extraction and classification algorithms are studied and simulated,including support vector machine(SVM),artificial neural network(ANN),principal component analysis(PCA)and deep belief network(DBN).In this paper,the advantages and disadvantages in dimension reduction of DBN and PCA are compared.Considering that DBN is an unsupervised training,this paper combines DBN and SOFTMAX,to conduct supervised training model,and extracts deep emotion features.Results shows that DDF features have excellent performance in speech emotion recognition.3.The paper has proposed a multilevel classification algorithm based on deep neural network.First,this paper has studied the shortcomings of traditional classification algorithms.By using the degree of confusion,a multilevel classifier is constructed,and the recognition rate is better than that of the traditional classifier.This paper compares the PCA-SVM multilevel classifier and PCA-SVM classifier,DDF-SVM multilevel classifier and DDF-SVM classifier,and the results are well improved,which proves the excellent performance of DDF-SVM multi classifier.Speech emotion recognition has the great practical significance.The paper has studied the feature extraction algorithm,feature learning algorithm and feature classification algorithm,and made research on discrete speech emotion recognition.The paper has proposed a new learning algorithm,and a new emotion classification algorithm to make improvement,and has achieved good results.