Improving emotion recognition using spectral and prosodic features

Abstract

Speech Emotion recognition is one of the key steps towards building an effective computing system. The reliability and accuracy of emotion recognition system greatly depend on the feature parametrization techniques. In this paper, prosodic features, such as pitch and intensity, and voice quality features, such as jitter ans shimmer, are used as representations for a front-end of Support Vector Machines (SVM) classifier for classifying speech into one of the seven discrete emotion classes: anger, boredom, happiness, disgust, fear, sadness, and neutral style. Spectral speech features, such as, Mel-frequency cepstral coefficients (MFCC), Perceptual Linear Prediction(PLP) and Rasta-PLP features, are used with Gaussian Mixture Model (GMM) and support vector machines (SVM), for the same task. The proposed features provide some complementary and supplementary emotion-specific information. Intelligent combination at the feature level and the score level of features is then expected to improve the intended performance of the system. Extensive computer simulations were conducted on the well-established Berlin emotional speech dataset (Emo-Db). The experimental results reveal 65 % accuracy using prosodic features and 88% for spectral features, whereas the combination of these two systems enhances the average accuracy level up to 89.33%.