Speaker-Dependent human emotion recognition in unimodal and bimodal scenarios

This paper presents an efficient technique of human emotion recognition using the audio and visual modalities for the speaker-dependent scenario. to achieve better emotion classification different audio and visual features were extracted. the feature selection was performed using the plus -take away algorithm based on two criteria: mahalanobis distance and kl-divergence. the feature selection was followed by feature reduction using pca and lda,and classification using the gaussian classifier. emotion classification was performed using both the unimodal and bimodal approaches. in the bimodal approach,audio and visual features were fused at two levels: feature and decision. the emotion classification performance comparable to humans was achieved on the savee database for the unimodal and bimodal scenarios. © 2015,pakistan academy of sciences.