An effective discriminative learning approach for emotion-specific features using deep neural networks

Abstract

Speech contains rich yet entangled information ranging from phonetic to emotional components. These different components are always mixed together hindering certain tasks from achieving better performance. Therefore, automatically learning a good representation that disentangles these components is non-trivial. In this paper, we propose a hierarchical method to extract utterance-level features from frame-level acoustic features using deep neural networks (DNNs). Moreover, inspired by recent progress in face recognition, we introduce centre loss as a complementary supervision signal to the traditional softmax loss to facilitate the intra-class compactness of the learned features. With the joint supervision of these two loss functions, we can train the DNNs to obtain separable and discriminative emotion-specific features. Experiments on CASIA corpus, Emo-DB corpus and SAVEE database show comparable results with that of state-of-the-art approaches.