Cross-Database Replay Detection in Terminal-Dependent Speaker Verification

The vulnerability of automatic speaker verification (ASV) systems against replay attacks becomes a severe problem. Although various methods have been proposed for replay detection, the generalization capability is still limited. For instance, a detection model trained on one database may fully fail when tested on another database. In this paper, we adopt the one-class learning technology to address the cross-database problem. Different from conventional two-class models that discriminate genuine speeches from replay attacks, the one-class model focuses on the within-class variance of genuine speeches, which is naturally robust to unseen attacks. In this study, we choose the Gaussian mixture model (GMM) as the one-class model and design two utterance-level features which reduce the uncertainties of genuine class while still be distinguishable from non-genuine class. Experiments conducted on three public replay datasets show that, compared to the state-of-the-art methods, the proposed method demonstrates promising generalization capability under cross-database scenarios.