Cascaded Subband Energy-Based Emotion Classification

Abstract

Since the earliest studies of human behavior, emotions have attracted attention of researchers in many disciplines, including psychology, neuroscience, and lately computer science. Speech is considered a salient conveyor of emotional cues, and can be used as an important source for emotional studies. Speech is modulated for different emotions by varying frequency- and energy-related acoustic parameters such as pitch, energy, and formants. In this paper, we explore analyzing inter- and intra-subband energy variations to differentiate six emotions. The emotions considered are anger, disgust, fear, happiness, neutral, and sadness. In this research, TwoLayered Cascaded Subband Cepstral Coeffcients (TLCS-CC) analysis was introduced to study energy variations within low and high arousal emotions as a novel approach for emotion classification. The new approach was compared with Mel Frequency Cepstral Coeffcients (MFCC) and Log Frequency Power Coeffcients (LFPC). Experiments were conducted on the Berlin Emotional Data Corpus (BEDC). With energy-related features, we could achieve average accuracy of 73.9% and 80.1% for speakerindependent and-dependent emotion classification respectively.