Improving Noise Robustness of Speech Emotion Recognition System.

In this paper method of improving noise robustness of speech emotion recognition system is proposed. Such a system has been developed for use in a social robot, but its performance is highly degraded by environmental noise. To alleviate this problem, the histogram equalisation is proposed to reduce the difference between feature vectors in clean and noisy conditions. In training phase of the system the averaged histograms of pitch and MFCC are computed and then serve as reference for equalisation. System performance was evaluated using Database of Polish Emotional Speech, which was split into training and test sets. Test sets were noised with 3 different noise samples. Presented preliminary results show a significant improvement of recognition accuracy in noisy environment conditions.