Estimation of emotional states enhanced by a priori knowledge

Abstract

This paper presents an improvement of conventional supervised-learning emotional state estimation in the form of dimensional valence-arousal values. In the proposed approach, outputs of the conventional estimator are additionally adapted using a priori knowledge about valence-arousal relations, which is extracted from the estimator's training set. Different approaches to a priori knowledge modeling have been undertaken: (a) single integral model over valence-arousal space, and (b) integration of multiple models that represent different discrete emotions in the valence-arousal space, specifically happiness, sadness, fear, anger and neutral state. This emotion estimation approach has been applied to conventional valence-arousal estimation from acoustic speech features based on support vector machines, using data from Croatian emotional speech corpus. Improvement of the results has been demonstrated.