Sample Reconstruction and Secondary Feature Selection in Noisy Speech Emotion Recognition

The research of noisy speech emotion recognition is significant in practical applications. In this paper, samples reconstructed by Compressed Sensing are used in noisy speech emotion recognition and a secondary feature selection fusing filter feature selection algorithms is proposed to achieve more effective feature subset. Three reconstruction methods are adopted on the measurements of noisy samples to demonstrate the feasibility of sample reconstruction in speech emotion recognition. The negative impact of noise and the effectiveness of secondary feature selection are verified on five emotions in Berlin Database of Emotional Speech. Experimental results show that the combination of sample reconstruction and secondary feature selection can improve the emotion recognition accuracies of noisy samples to be close to or even higher than the accuracies of clean samples.