Improved speech emotion classification from spectral coefficient optimization

Abstract

In order to improve the performance of speech emotion recognition systems, and to reduce the related computing complexity, this work proposed two approaches of spectral coefficient optimization. The two approaches are (1) optimized based on discrete spectral features and (1) combine spectral features. Experimental studies have been performed through the Berlin Emotional Database, using a support vector machine (SVM) classifier, and five spectral features including MFCC, LPC, LPCC, PLP and RASTA-PLP. The experiment results have shown that speech emotion recognition based on optimized coefficient numbers can effectively improve the performance. There were significant improvements in the accuracy 2 % for the first approach and 4 % for the second approach compared to that using the existing approaches. Moreover the second approach outperformed the first approach in the accuracy. This good accuracy came with reducing the features number.