Comparison of Classifiers for Speech Emotion Recognition (SER) with Discriminative Spectral Features

Abstract

This paper compares the classification ability of a few efficient classifiers in recognizing human speech emotions in terms of accuracy, computation time, and feature dimension. Both the static and dynamic mel-frequency cepstral coefficients (MFCC) are derived in the wavelet domain and are combined to form a suitable identification system model. Three popular neural network (NN) models such as the Gaussian mixture model (GMM), radial basis function network (RBFN), and the probabilistic neural network (PNN) have been put to test the reliability of these derived feature sets. The PNN classifier has shown to outperform both the RBFN and the GMM with low feature dimension, whereas the GMM shows an improved result for large feature dimensions. The combination of wavelet-based MFCCs and their dynamics remains more discriminative in classifying speech emotions as compared to either the MFCCs or wavelets acted alone.