A pairwise discriminative task for speech emotion recognition

Speech emotion recognition is an important task in human-machine interaction. However, it faces many challenges such as the ambiguity of emotion expression and the lack of training samples. To solve these problems, we propose a novel 'Pairwise discriminative task', which attempts to learn the similarity and distinction between two audios rather than specific labels. In the task, pairwise audios are fed into audio encode networks to extract audio vectors, followed with discrimination networks behind to judge whether audios belong to the same emotion category or not. The system is optimized in an end-to-end manner to minimize the loss function, which cooperates cosine similarity loss and cross entropy loss together. To verify the performance of audio representation vectors extracted from the system, we test them on IEMOCAP database-a common evaluation corpus. We gain 56.33% unweighted accuracy on the test database, which surpasses above 5% compared with traditional end-to-end speech emotion recognition networks.