Considering relative order of emotional degree in dimensional speech emotion recognition

Dimensional speech emotion recognition (Dim-SER) is a rising branch of emotion computing field. It views emotion from dimensional and continuous perspective, and formalizes the SER problem as a regression task. Current Dim-SER researches never consider the relative order of emotional degree between utterances, which can make the human-machine interface get wrong information about speaker’s emotion variation trend. Starting from this demand, this paper constructs a relative order of emotion degree sensitive Dim-SER system with the human emotion cognitive characteristics as reference, and employs Gamma statistic to evaluate emotion recognition performance. Specifically, the Top-rank probability distribution is developed to describe the emotional ordering of utterances, and the Kullback-Leibler divergence is used to measure the loss of order consistency caused by emotion recognition. Finally, the Order-Senstive Network (OSNet) algorithm is proposed to minimized prediction loss. Experimental results show that, compared with the commonly used k-Nearest Neighbor (k-NN) and Support Vector Regression (SVR) approaches, the proposed system effectively improve the correctness of emotional relative order between utterances.