Feature analysis of speech emotion data on arousal-valence dimension using adaptive neuro-fuzzy classifier

Speech is structured acoustic signals which form a message featuring the speaker´s language, speaking style, and also underlying emotion. These features affect the information passed through speech. Automated speech emotion recognition is a field long studied, but it has not yet found a quite reliable approach. This paper explores two ways to improve automated recognition. First, new sets or combinations of speech emotion features can be selected. Second, recognition of the new feature sets can be separated into arousal and valence dimensions to identify the weaker dimension, which is not possible if trying to recognize emotions directly. The Adaptive Neuro-Fuzzy Classifier (ANFC) is used as classifier and feature selector, and Self Organizing Map (SOM) is used to visualize the behavior of sample data based on the selected features.