Hierarchical Classification of Emotional Speech

Speech emotion as anger, boredom, fear, gladness, etc. is high semantic information and its automatic analysis may have many applications such as smart human-computer interactions or multimedia indexing. Main difficulties for an efficient speech emotion classification reside in complex emotional class borders leading to necessity of appropriate audio feature selection. While current work in the literature only relies on classical frequency and energy based features and make use of a global classifier with a identical feature set for different emotion classes, we propose in this paper some new harmonic and Zipf based features for better emotion class characterization and a hierarchical classification scheme as we discovered that different emotional classes need different feature set for a better discrimination. Experimented on Berlin dataset [11] with 68 features, our emotion classifier reaches a classification rate of 76.22% and up to 79.47% when a first gender classification is applied, whereas current works in the literature usually display, as far as we know, a classification rate from 55% to 70%.