MFCC global features selection in improving speech emotion recognition rate

Abstract

Feature selection is one of the important aspects that contribute most to the emotion recognition system performance apart from the database and the classification technique used. Based on the previous finding, Mel Frequency Cepstral Coefficients (MFCC) are said to be good for emotion recognition purpose. This paper discusses the use of MFCC features to recognize human emotion on Berlin database in the German language. Global features are extracted from MFCC and tested with three classification methods; Naive Bayes, Artificial Neural Network (ANN) and Support Vector Machine (SVM). We investigate the capabilities of MFCC global features using 13, 26 and 39-dimensional cepstral features in recognizing emotions from speech. The result from the experiment will be further discussed in this paper.