| Speech emotion recognition is to automatically detect the speaker’s emotion stateaccording to speech using machine learning methods, which mainly involves featureextraction and classification algorithms. Speech emotion recognition can be widely used ineducation, entertainment, medical treatment and so on.In this paper, we propose a model space parameter named GSWM, and a recognitionmethod with the fusion of multiple classifiers, due to unsatisfied recognition rate withsingle group of features or single classifier.In feature extraction, we propose a new feature group SWMFCC, combining LSFwith good interpolation and quantization performance and MFCC which presents humanauditory characteristics, to get bidirectional presentings from both the expression of thespeaker and the emotional perception of the listener. Then, GMM model was applied to itto obtain model space parameter GSWM with detailed information, in order to furtherimprove the emotion recognition performance.In classification, we propose a recognition method using fusion of multiple classifiersbased on D-S (Dempster-Shafer’s) evidence theory. The results of several classifiers arefused to better deal with the uncertainty of the classifications. SVM (Support VectorMachine) is selected in our work as basic classifiers. Recognition experiments withprosodic features, voice quality features and GSWM features which present emotiondetails are carried out, and the classification results are fused based on D-S evidencetheory.Experiments with EMO-DB were taken out to validate the effectiveness of GSWM and fusion algorithm of multiple classifiers, on six emotions as anger, happiness, neutral,sadness, fear and boredom. The experiment results show that the SWMFCC and GSWMfeatures can effectively convey emotional information in speech, improve the performancein the emotion recognition. The correct classification rate can reach up to83.75%withGSWM when using SVM, and can be further improved to90.5%with D-S based multipleclassifiers fusion. |
PDF Full Text Request