Excitation source and low level descriptor features fusion for emotion recognition using SVM and ANN

Emotion recognition is a challenging task with many applications in healthcare and human-machine interaction. In this study we propose to fuse two sets of features for emotion recognition at the classification level. A set of features that includes LPCC and MFCC extracted from LP-residual samples and Wavelet Octave Coefficient Of Residual (WOCOR) is proposed in this study as excitation source features. The proposed set of features is fused with 6552 LLDs using SVM and ANN classifiers. The experiments are tested on a newly acquired emotional speech database in Kurdish language, the Berlin emotional speech database, and the Aibo database. The experiments demonstrate that the fusion of the proposed excitation source features with the common LLDs can achieve better recognition accuracies than what is reported in the state-of-the-art studies.