Relative Entropy Normalized Gaussian Supervector for Speech Emotion Recognition using Kernel Extreme Learning Machine

Speech emotion recognition is a challenging and significant task. On the one hand, the emotion features need to be robust enough to capture the emotion information, and while on the other, machine learning algorithms need to be insensitive to model the utterance. In this paper, we presented a novel framework of speech emotion recognition to address the two above-mentioned challenges. Relative Entropy based Normalization (REN) was proposed to normalize the supervectors of Gaussian Mixture Model-Universal Background Model (GMM-UBM) as the features to emotions. The Kernel Extreme Learning Machine (KELM) was adopted as the classifier to identify the emotion represented by the normalized supervectors. Experimental results on the EMR 1309 corpus showed the proposed framework outperformed the state-of-the-art i-vector based systems.