SVM-MLP-PNN Classifiers on Speech Emotion Recognition Field - A Comparative Study

Abstract

In this paper, we present a comparative analysisof three classifiers for speech signal emotion recognition.Recognition was performed on emotional Berlin Database.This work focuses on speaker and utterance (phrase)dependent and independent framework. One hundred thirtythree (133) sound/speech features were extracted from Pitch,Mel Frequency Cepstral Coefficients, Energy and Formantsand were evaluated in order to create a feature set sufficient todiscriminate between seven emotions in acted speech. A set of26 features was selected by statistical method and MultilayerPercepton, Probabilistic Neural Networks and Support VectorMachine were used for the Emotion Classification at sevenclasses: anger, happiness, anxiety/fear, sadness, boredom,disgust and neutral. In speaker dependent framework,Probabilistic Neural Network classifier reached very highaccuracy of 94%, whereas in speaker independent framework,Support Vector Machine classification reached the bestaccuracy of 80%. The results of numerical experiments aregiven and discussed in the paper.