Adversarial Domain Adaptation for Noisy Speech Emotion Recognition

Speech Emotion Recognition (SER) has achieved many great results with deep learning techniques. However, noise discrepancy is still a challenging task due to the distribution shift between training and test data. In this paper, we present a novel approach based on unsupervised domain adaptation method to alleviate the distribution shift problem for noisy SER. Specifically, we apply an adversarial domain adaptation with bridge mechanism to model an intermediate domain for knowledge transfer. We construct a bridge layer by exploiting speech denoising approach to extract the domain-specific noise representation. Experimental results show that our method provides average improvements of 2.29% and 3.88% in weighted and unweighted accuracies over the baseline for SER with various noise settings.