Speech Emotion Recognition Based on Robust Discriminative Sparse Regression

Abstract

Speech emotion recognition has recently attracted much interest due to the widespread of multimedia data. It generally involves two basic problems: 1) feature extraction and 2) emotion classification. Most previous algorithms just focus on solving one of these two problems. In this article, we aim to deal with these two problems in a joint learning framework, and present a novel regression algorithm, namely, robust discriminative sparse regression (RDSR). In RDSR, we propose a sparse regression algorithm to make our model robust to outliers and noises, and introduce a feature selection regularization constraint simultaneously to select the most discriminative and relevant features. In addition, to well predict the labels, we exploit the local and global consistency over labels, and incorporate it into the proposed framework. To solve the objective function of RDSR, we design an efficient alternative optimization algorithm. Finally, experimental results on several public emotion data sets verify the effectiveness and the superiority of our proposed method.