Comparison of Neural Network Models for Speech Emotion Recognition

Abstract

This work compares the performance of three prominent Neural Network based classifiers for speech emotion recognition (SER). The classifiers such as the Multilayer Perceptron (MLP), the Radial Basis Function Network (RBFN) and the Probabilistic Neural Network (PNN) have been tested for their effectiveness in recognizing speech emotions such as angry, sad, bore and happy states from Berlin (EMO-DB) database. The self-learning ability of these classifiers to capture complex input and output relationship effectively makes them versatile in the field of SER. Hence, these parallel structured networks are expected to provide the intended accuracy for the proposed task as speech frequencies occur in parallel. Extensive simulation of these classifiers has been carried out using the popular Vector Quantized based Linear Predictive Cepstral Coefficients (LPCC_VQ) and the excitation source Hurst components. The PNN has shown to outperform all others in classifying the chosen emotional states using the extracted feature sets. The LPCC_VQ remains more discriminating for the proposed SER system than the pH as our results reveal. The average classification error using the LPCC-VQ feature sets has been 0.173 for the PNN as compared to 0.185 with RBFN and 0.227 with MLP as observed from our results.