Speech Emotion Recognition Using Adaptive Ensemble of Class Specific Classifiers

Emotion recognition plays a significant role in Human Computer Interaction (HCI) field for effective communication. The aim of this study is to built a generic emotion recognition system to face the challenges of recognition in resolving confusion among acoustical characteristics of emotions, identifying dominating emotion from mixed emotions etc. When there is confusion among the perception of emotion by human, the understanding of it by machine is a real challenge. Due to these reasons, it is very hard to produce highly accurate emotion recognition system in real time. Researchers are working to improve the performance of emotion recognition task by designing different classifiers and also using different ensemble methodologies at data level, feature level or decision levels to recognize emotion. We have built a generic SVM based emotion recognition system, which models emotions using given features. Out of given acoustical features, for every emotional class, class specific best features are identified based on f1 measure. The responses of the systems, built based on these best features are combined using new smart additive ensemble techniques. Decision logic is employed to decode the responses into an emotional class, the class which produces maximum value among all emotional classes. A rejection framework is also designed to reject a noisy and weak input file. We have tested the framework with 12 acoustical features on Berlin emotional corpus EMO-DB. The accuracy obtained from our generic emotion recognition system is 74.70% which is better than classifiers reported in the literature.