A feature extraction scheme based on enhanced wavelet coefficients for Speech Emotion Recognition

This paper proposes a new feature extraction scheme for speaker-independent Speech Emotion Recognition under both unsupervised and supervised conditions following a non-hierarchical process first and then a hierarchical one. The feature is derived from the Teager energy operated wavelet coefficients of speech signal. The wavelet coefficients enhanced by TE operation is used to compute entropy thus forming a feature vector. The feature vector is fed to unsupervised K-means clustering in a nonhierarchical process. Considering the effectiveness of supervised classification in a recognition problem, the feature is then used in a supervised KNN classifier. It is seen that supervised KNN classifier is more capable of distinguishing different emotions when a hierarchical approach is followed for recognition instead of a non-hierarchical one. Detail simulations are carried on EMO-DB German speech emotion database containing four class emotions, such as angry, happy, sad and neutral. Simulation results show that the proposed feature with supervised hierarchical classification approach provides the higher accuracy in comparison to the other methods of four-class emotion recognition.