Acoustic emotion recognition using spectral and temporal features

Abstract

In this paper, utility of different low- level, spectral and temporal features is evaluated for the task of emotion recognition. The aim of an ideal speech emotion recognition system is to extract features that are representative of the emotional state of speaker. Pitch, intensity, frequency formants, jitter, and zero crossing rate are five features proposed for characterizing four different emotions, anger, happy, sadness, and neutral. Low- level spectral and temporal features have ease of calculation and limit the complexity of emotion recognition systems since they are commonly single dimensional features. A decision-tree based algorithm is designed for characterizing emotions using these acoustic features. It has been proven that various aspects of a speaker’s physical and emotional state can be identified by speech alone. However, the accuracy of such analyses has not been optimized due to acoustic variabilities such as length and complexity of human speech utterance, gender, speaking styles, and speech rate. Since speech emotion recognition is a developing and challenging field, most powerful features for emotion recognition are not yet defined; hence, investigating the utility of selected features for emotion recognition is an important task.