Speech emotion classification via a modified Gaussian Mixture Model approach

Emotional state of the speaker is an important feature embedded in his/her produced speech signal. Despite emotion recognition importance in system performance improvement, such as in ASR, not much research has been carried out in the speech emotion classification field. This paper is focused on finding more effective approaches to improve speaker emotional state classification methods. Two approaches are proposed for training and test phases while the Gaussian Mixture Model (GMM) is selected as the classifier. In these approaches, the motivation is to reduce the confusing information regions of emotional speech space and to increase salience of the discriminative regions. In the training phase, symmetric Kullback-Leibler Divergence (KLD) is used as a measure to detect the discriminative GMM mixtures while the confusing mixtures are ignored. This algorithm is known as KLD-GMM. In the test phase, the discriminative frames are recognized based on Frame Selection Decoding (FSD). This algorithm is known as FSD-GMM, when FSD algorithm is applied on KLD-GMM algorithm, the approach is named KLD-FSD-GMM algorithm. Two proposed algorithms have led to an average absolute improvement of about 7% in the emotion recognition performance in comparison with the baseline generalized GMM-based method.