Research On Speech Emotion Recognition Based On Deep Features Fusion And Joint Decision

Speech signals contain a variety of speech information and a variety of rich emotional states.In order to make human-computer interaction more efficient,it is of great application value and research significance to study the emotional features contained in speech signals and understand their emotional elements.In order to improve the speech emotion recognition rate,this paper starts with the selection of emotion features and the construction of classification model.On the one hand,the bottleneck features which can effectively express speech emotion are selected;on the other hand,the recognition models are improved,which applied in speech emotion recognition respectively.The main research contents and innovation points of this paper are as follows:(1)Firstly,this paper briefly describes the background of speech emotion recognition and the research status at home and abroad,and summarizes the system framework of speech emotion recognition.Then,various speech emotion description models and their corresponding corpus,various speech signal preprocessing techniques and common speech emotion features are introduced in detail.In this paper,pure and noiseless EMO-DB database and CASIA database based on discrete model are selected as experimental corpus to evaluate the performance of each module of speech emotion recognition system based on DNN,find out the existing problems and provide solutions.(2)The bottleneck features based on DNN have been widely used in the field of speech emotion recognition.However,a single bottleneck feature cannot fully describe the emotional information of speech.This paper proposes a speech emotion recognition method based on two-stage bottleneck features selection to extract more suitable emotional features for classification.Firstly,a DNN model is constructed.By setting the position of the bottleneck layer in the hidden layer of DNN,different bottleneck features are extracted,and the statistical variables are calculated to construct the deep and shallow bottleneck features respectively.Then,in order to highlight the different contribution degrees of different features to emotion classification,GA is used to optimize the contribution weights of the features,and the weight set obtained from the optimization is used to realize the fusion of deep and shallow bottleneck features with weights.PCA is used for feature screening,and SVM model is used for classification.The experimental results show that the method based on two-stage bottleneck feature selection can effectively improve the performance of speech emotion recognition.(3)A single classifier does not necessarily have a good recognition effect for every emotion.Therefore,based on the construction of basic classifiers,a new combined classifier is formed by considering the prediction results of multiple basic classifiers.This paper proposes a speech emotion recognition method based on MCJD algorithm.In the construction of basic classifiers,this paper takes the overall good performance of a variety of emotions and the good performance of a target emotion as a goal to form a number of basic classifiers,and then get the final decision result through MCJD algorithm.The experimental results show that the speech emotion recognition method based on MCJD algorithm can improve the effect of speech emotion recognition,and has a good correction effect on the decision results of traditional SVM classifier.