Abstract

Emotion speech causes additional information added with respect to writing information. From the other hand, presence of state in speech also causes some problem in recognition process of speech. Feature selection is a problem of choosing a subset of relevant features. Researchers have been searching for optimal feature selection methods and finding informative features and combining powerful classifiers that improve the performance of emotion recognition systems in different applications. In this paper, SVM method has been used as classification of induced features, then for ranking of features, three feature ranking methods are used for speech emotion recognition. These methods are Fisher score (FS), Mutual Information (MI), and Area Under Curve maximizing SVM(AUCS). For this purpose, a rich feature set with the size of 55 is used. Then five distinct feature subsets with the size of 16,32,39,45 and 55 are selected from the mentioned feature set. To investigate the performance of system with small-size input feature set, eight high-ranked features are selected from each of mentioned feature sets (with the size of 16,32,39,45 and 55) and two types of neural networks (multi-layer perceptron (MLP) and radial basis function (RBF)) and Gaussian mixture model (GMM) method are used for emotion recognition. Experimental results show that using AUCS based feature ranking method and GMM recognizer result in emotion recognition rates above 82 by employing a small-size feature set.