Cross-corpus speech emotion recognition using semi-supervised transfer non-negative matrix factorization with adaptation regularization

This paper focuses on a cross-corpus speech emotion recognition (SER) task, in which there are some mismatches between the training corpus and the testing corpus. Meanwhile, the label information of the training corpus is known, while the label information of the testing corpus is entirely unknown. To alleviate the influence of these mismatches on the recognition system under this setting, we present a non-negative matrix factorization (NMF) based cross-corpus speech emotion recognition method, called semi-supervised adaptation regularized transfer NMF (SATNMF). The core idea of SATNMF is to incorporate the label information of training corpus into NMF, and seek a latent low-rank feature space, in which the marginal and conditional distribution differences between the two corpora can be minimized simultaneously. Specifically, in this induced feature space, the maximum mean discrepancy (MMD) criterion is used to measure the discrepancies of not only two corpora, but also each class within the two corpora. Moreover, to further exploit the knowledge of the marginal distributions, their underlying manifold structure is considered by using the manifold regularization. Experiments on four popular emotional corpora show that the proposed method achieves better recognition accuracies than state-of-the-art methods.