Dimensionality Reduction for Emotional Speech Recognition

Abstract

The number of speech features that are introduced to emotional speech recognition exceeds some thousands and this makes dimensionality reduction an inevitable part of an emotional speech recognition system. The elastic net, the greedy feature selection, and the supervised principal component analysis are three recently developed dimensionality reduction algorithms that we have considered their application to tackle this issue. Together with PCA, these four methods include both supervised and unsupervised, as well as filter and projection-type dimensionality reduction methods. For experimental reasons, we have chosen VAM corpus. We have extracted two sets of features and have investigated the efficiency of the application of the four dimensionality reduction methods to the combination of the two sets, besides each of the two. The experimental results of this study show that in spite of a dimensionality reduction stage, a longer vector of speech features does not necessarily result in a more accurate prediction of emotion.