Research On Speech Emotion Recognition Based On Feature And Decision Fusion

With the continuous development of artificial intelligence,users' demand for human-computer interaction interface is higher and higher.Human beings hope that computers can perceive people's emotional changes and make the communication between people and computers more unhindered.As the key technology of humancomputer interaction,speech emotion recognition has attracted extensive attention of researchers and has been applied in the fields of distance education,medical diagnosis and traffic driving.Therefore,the research of speech emotion recognition has important research significance and practical value.Aiming at the problem of poor recognition effect of single speech emotion feature,a speech emotion recognition algorithm based on MFCC and IMFE feature fusion is proposed.In this paper,the affective feature parameter selection and fusion method is used.Firstly,the speech affective feature parameter evaluation method of variance ratio measure and separability measure is adopted.By judging the separability and correlation between various dimensional features,the best part of MFCC and IMFE are fused,and a new speech affective fusion feature MFCC-IMFE is obtained.Finally,KELM classifier is used to classify speech emotion.Through the experiment on the IEMOCAP database,the recognition accuracy of 90.25% is obtained.Compared with the algorithm based on multi-core learning fusion audio features with the highest recognition rate on the IEMOCAP database,the algorithm in this paper is improved by 5.73%,which is prepared for the following research on the dual fusion algorithm.Aiming at the weak performance of speech emotion recognition with single feature and single classifier,by improving the traditional fusion algorithm,a decision fusion algorithm is added on the basis of the above research on feature fusion algorithm,and a speech emotion recognition algorithm based on dual fusion of feature and decision is proposed,which retains the difference and correlation between various emotional information.Firstly,the MFCC features are input into the convolution recurrent neural network,combined with the multi head attention mechanism,many subspaces are divided for learning,and different weights are given to the emotional features,and then the emotion is classified by KELM classifier;Then two single classifiers are trained to convert their numerical output into probability output;Next,according to the decision-making strategy,the adaptive weight of the test sample is obtained;Finally,the probability of the output is linearly weighted and the output is judged.Through experiments on IEMOCAP database,the recognition accuracy of speech emotion can reach 96.5%.Compared with the fusion algorithm based on fuzzy cognitive map,which has the highest recognition rate,this algorithm improves the recognition accuracy by 7.07%.