Emotion Recognition in Speech with Latent Discriminative Representations Learning

Despite significant recent advances in the field of affective computing, learning meaningful representations for emotion recognition remains quite challenging. In this paper, we propose a novel feature learning approach named Latent Discriminative Representation (LDR) learning for speech emotion recognition. Unlike most existing hand-crafted features designed for specific applications or features learnt by a standard neural network, the proposed learning method incorporates an additional training objective in order to learn better representations of the task of interest. To this end, we group the training samples into sets of triplets, satisfying that the second member in each triplet comes from the same class as the first and that the third member comes from a different class than the first. In the training process, we maximise the distance of the samples from different classes in the latent representation space, while we minimise the distance for samples from the same class. To evaluate the effectiveness of LDR, we perform extensive experiments on the widely used database IEMOCAP, and find that the LDR improves performance over the standard neural network training procedure. (C) 2018 The Author(s). Published by S. Hirzel Verlag . EAA.