Transferable discriminant linear regression for cross-corpus speech emotion recognition

Abstract

Speech emotion recognition (SER) has attracted much interest recently due to its wide applications. However, it should be noted that most SER methods are conducted on the assumption that the training and testing data are from the same database. In real applications, this assumption does not hold, and the recognition performance will be significantly degraded. To solve this problem, we present a novel transferable discriminant linear regression (TDLR) approach for cross-corpus SER. Specifically, first, we introduce a non-negative label relaxation linear regression on source corpus to help learn transferable feature representations. Second, we propose a simple but effective strategy to keep the linear relationship between the labels of source and target corpora. Meanwhile, we utilize the discriminative maximum mean discrepancy (MMD) as the distance metric between two databases. Furthermore, we use the graph Laplacian to preserve the geometric structure of samples, which can further reduce the distribution gap between the two databases. Additionally, to better obtain the intrinsic properties of data and make the model robust, we impose an ℓ2,1-norm on the transformation matrices. Extensive experiments have been carried out on several standard databases, and the results show that TDLR can obtain better recognition performance than several state-of-the-art algorithms.