Comparison of Different Classifiers for Emotion Recognition

Abstract

In the present paper a comparison of two classifiers for speech signal emotion recognition is presented. Recognition was performed on emotional Berlin Database. Within this work we concentrate on the evaluation of a speaker-dependent and speaker independent emotion recognition classification. One hundred thirty three (133) speech features obtained from speech signal processing. A basic set of 35 features was selected by statistical method and artificial neural network and random forest classifiers were used. Seven classes were categorized, namely anger, happiness, anxiety/fear, sadness, boredom, disgust and neutral. In speaker dependent framework, artificial neural network classification reached an accuracy of 83,17%, and random forest 77,19%. In speaker independent framework, for artificial neural network classification a mean accuracy of 55% was reached, while random forest reached a mean accuracy of 48%.