SPEECH EMOTION CLASSIFICATION USING SVM AND MLP ON PROSODIC AND VOICE QUALITY FEATURES

Abstract

In this paper, a comparison of emotion classification undertaken by the Support Vector Machine (SVM) and the Multi-Layer Perceptron (MLP) Neural Network, using prosodic and voice quality features extracted from the Berlin Emotional Database, is reported. The features were extracted using PRAAT tools, while the WEKA tool was used for classification. Different parameters were set up for both SVM and MLP, which are used to obtain an optimized emotion classification. The results show that MLP overcomes SVM in overall emotion classification performance. Nevertheless, the training for SVM was much faster when compared to MLP. The overall accuracy was 76.82% for SVM and 78.69% for MLP. Sadness was the emotion most recognized by MLP, with accuracy of 89.0%, while anger was the emotion most recognized by SVM, with accuracy of 87.4%. The most confusing emotions using MLP classification were happiness and fear, while for SVM, the most confusing emotions were disgust and fear.