Research On Prediction Of Emotional Dimensions PAD For Speech Emotion Recognition

Affective computing is an important branch in the field of artificial intelligence.In order to realize natural human-computer interaction,objective understanding and description of human emotion is essential.Previous affective computing is mainly focused on the recognition of discrete emotion,and the mapping relationship between discrete emotion and feature is studied by computer,so that emotion is recognized as label discrete emotion.This method of affective computing is easy to ignore the complexity of emotion.In recent years,emotional dimension has attracted wide attention of researchers at home and abroad because of its advantages of describing emotions from a psychological perspective and quantifying any emotion.Therefore,this thesis takes emotional speech as the research object,starts from emotion feature,selects the emotion feature with high correlation with emotional dimension by analyzing the relationship between emotion feature and emotional dimension,and improves the regression model.The predicted values of PAD(Pleasure,Arousal and Dominance,PAD)dimensions,which are more in line with the mechanism of human emotion processing,are obtained.The predicted values of dimensions are used for emotional recognition.The experimental results show that the dimension values has certain advantage in distinguishing emotionalstates.The main research contents of this thesis are as follows:(1)The relationship between emotion feature and emotional dimension PAD was analyzed,and the GRA-PCA method was applied to the dimension reduction of emotion feature in PAD prediction.The correlation degree between different feature and PAD is calculated by using GRA(Grey Relation Analysis,GRA)algorithm,and the features with low correlation are removed to reduce the redundancy of feature.PCA(Principal Component Analysis,PCA)algorithm is used to extract features,which reduces the correlation between features.The GRA-PCA feature dimension reduction method was applied to PAD dimensions prediction.The experimental results show that this method can effectively improve the prediction accuracy of PAD.(2)A Clustering PSO-LSSVM regression model was proposed by analyzing the deficiencies of regression model in PAD dimensions prediction.Particle Swarm Optimization(PSO)algorithm with the global search ability is adopted to optimize the model parameters of LSSVM,which weakens the blindness of model parameter selection.The PSO-LSSVM model was improved by emotional clustering analysis,so as to reduce the interaction between different emotion.A Clustering PSO-LSSVM regression model was established.The model was used to predict PAD,and the experimental results show that the prediction effect of the model is improved.(3)A GRA-PCA-clustering PSO-LSSVM regression model was proposed based on the fusion of GRA-PCA feature dimension reduction method andClustering PSO-LSSVM model.The model was used to predict the PAD dimensions,and the experimental results show that the model has better predictive ability for PAD than GRA-PCA-LSSVM or Clustering PSO-LSSVM.In order to further explain the accuracy of regression model in predicting PAD,three groups of emotion recognition experiments were designed and compared.The first group used FPFMN(Fusion Feature of Prosodic Feature,Formant,MFCC Feature and Nonlinear Feature,FPFMN)features to identify emotion.The second group used PAD predicted by three kinds of models(GRA-PCA-LSSVM,Clustering PSO-LSSVM,GRA-PCA-clustering PSO-LSSVM model)as features to identify emotion.The third group used the fusion of PAD values predicted by three kinds of models with FPFMN features,respectively,and the fused features were applied to emotion recognition experiment.The experimental results show that in the three experiments,the fusion of PAD predicted by GRA-PCA-clustering PSO-LSSVM model and FPFMN features has the best emotion recognition effect.