Passive versus active: Vocal classification system

Five expressions are commonly considered to characterize human emotional states: Happiness, Surprise, Anger, Sadness and Neutral. Different measures can be extracted from speech signals to characterize these expressions, for example the pitch, the energy, the SPI and the speech rate. Automatic classification of the five expressions based on these features shows a great confusion between Anger, Surprise and Happiness on the one hand and Neutral and Sadness on the other hand. Such a confusion is also observed when humans make the same classification. We propose to define two classes of expression: Active gathering Happiness, Surprise and Anger versus Passive gathering Neutral and Sadness. Such a partition is also better suited for the integration of speech information in a multimodal classification system based on speech and video, which is the long term aim of our work. In this paper, we test several classification methods, namely a Bayesian classifier, a Linear Discriminant Analysis (LDA), the K Nearest Neighbours (KNN) and a Support Vector Machine with gaussian radial basis function kernel (SVM). For the considered two classes, the best performances are achieved with the SVM classifier with a recognition rate of 89.74% for Active state and of 86.54 % for Passive state.