High-order similarity learning based domain adaptation for speech emotion recognition

Abstract

Speech emotion recognition (SER) has received significant attention due to the advancement of artificial intelligence technology. Conventional SER methods usually assume that both the training and test data are derived from the same dataset, without fully considering the differences between different datasets, which would lead to reduced recognition performance. To address this problem, this paper proposes a novel domain adaptation approach called high-order similarity learning based domain adaptation (HSDA) for SER. Specifically, we first project the original data into a low-dimensional embedding subspace, which can effectively eliminate the inter-domain differences. Then, we learn the high-order similarity graph to exploit the intrinsic structural information of cross-domain data. At the same time, we utilize the regression term to enhance the discriminative power of the model, which can fully use the labeling information of the source domain to make the learned transformation matrix more discriminative. The experimental results on four popular datasets show that our method can achieve excellent performance compared to several state-of-the-art methods.