Adaptive Emotion Recognition in Speech by Feature Selection Based on KL-divergence

Abstract

This paper proposes adaptive emotion recognition system in speech by feature selection based on KL-divergence. In order that the system can choose the most suitable feature set for emotion recognition, we propose an evaluation method for the set of prosodic features based on Kullback-Leibler divergence (KL-divergence). Additionally, we propose a method of feature selection system, using genetic algorithm (GA) making use of rapid evaluation based on KL-divergence. Experimental results show the proposed system can acquire efficient the prosodic feature set for emotion recognition in short order without constructing a recognition system. Furthermore, the accuracy of emotion recognition is significantly improved with the prosodic feature set selected by the proposed system.