Speech emotion recognition using semi-NMF feature optimization

In recent times, much research is progressing forward in the field of speech emotion recognition (SER). Many SER systems have been developed by combining different speech features to improve their performances. As a result, the complexity of the classifier increases to train this huge feature set. Additionally, some of the features could be irrelevant in emotion detection and this leads to a decrease in the emotion recognition accuracy. To overcome this drawback, feature optimization can be performed on the feature sets to obtain the most desirable emotional feature set before classifying the features. In this paper, semi-nonnegative matrix factorization (semi-NMF) with singular value decomposition (SVD) initialization is used to optimize the speech features. The speech features considered in this work are mel-frequency cepstral coefficients, linear prediction cepstral coefficients, and Teager energy operator-autocorrelation (TEO-AutoCorr). This work uses k-nearest neighborhood and support vector machine (SVM) for the classification of emotions with a 5-fold cross-validation scheme. The datasets considered for the performance analysis are EMO-DB and IEMOCAP. The performance of the proposed SER system using semi-NMF is validated in terms of classification accuracy. The results emphasize that the accuracy of the proposed SER system is improved remarkably upon using the semi-NMF algorithm for optimizing the feature sets compared to the baseline SER system without optimization.