Entropy Based Boundary-Eliminated Pseudo-Inverse Linear Discriminant for Speech Emotion Recognition

摘要

Remarkable advances have achieved in speech emotion recognition (SER) with efficient and feasible models. These studies focus on the ability of the model itself. However, they ignore the potential distributed information of speech data. Actually, emotion speech is imbal-anced due to the expression of human being. To overcome the imbal-anced problems of speech data, the ongoing work furthers our previous study of the Boundary-Eliminated Pseudo-Inverse Linear Discriminant (BEPILD) model through introducing the information entropy that contributes to describing the distribution of the speech data. As a result, an Entropy-based Boundary-Eliminated Pseudo-Inverse Linear Discriminant model (EBEPILD) is proposed to generate more robust hyperplanes to tackle the speech data with high class uncertainty. The experiments conducted on the Interactive Emotional Dyadic Motion Capture (IEMO-CAP) database with four emotion states show that the EBEPILD has outstanding performance compared with other algorithms.