Enhancing Out-of-Domain Detection for Speech Spoofing Countermeasure via Supervised Contrastive Learning

High-performance anti-spoofing countermeasures (CMs) have been widely used to protect automatic speaker verification systems by identifying and filtering spoofing speech. However, their performance degrades severely when confronted with out-of-domain (OOD) samples. To solve this issue, recent papers investigate the strategy that a CM can opt for abstention when it is not confident about the decision. In this paper, we develop an effective approach to enhance the performance of CMs with abstention. Specifically, we introduce supervised contrastive learning to group samples into known classes more tightly and employ data augmentation to enhance the diversity of known samples. The experiments are conducted on ASV spoof 2019 logical access corpus and another test set consisting of different OOD samples from other databases. With our scheme, CMs with abstention can achieve an equal error rate (EER) of 1.84 % on the LA test set and 0.86 % on the other test set. These results demonstrate that CMs trained with our approach show better OOD detection performance and can make more confident decisions.