A Hierarchical Speech Emotion Classification Framework based on Joint Triplet-Center Loss

Automatic speech emotion recognition task is crucial to the development of human-computer interaction systems. However, the ambiguity of emotion categories and the subjectivity of human annotations make it hard to extract discriminative emotional features and improve the classification accuracy. In this paper, we propose a Joint Triplet-Center Loss based hierarchical learning method. On the one hand, the proposed Joint Triplet-Center Loss function can learn discriminative emotional features through reducing the intra-class distance and increasing the inter-class distance. On the other hand, the hierarchical learning method can enhance the stability of the model by considering the consistency of annotations. The experimental results show that our proposed method has obvious performance improvement compared with previous works, and gets better generalization performance.