Abstract
The high level of usability achieved by voice biometrics compared to other biometric authentication modalities has promoted the widespread use of automatic speaker verification (ASV) systems as authentication tools for several services in various domains. Despite their satisfactory performance, ASV systems are vulnerable to malicious voice spoofing attacks. Hence, voice spoofing countermeasures have emerged as essential solutions to stop such harmful attacks and protect ASV systems as well as users' confidentiality. Typically, these countermeasures classify utterances into genuine and spoofing categories. In this research, we propose two voice spoofing countermeasures that mainly aim to improve the generalization of supervised learning models. This goal is achieved through the adaptive handling of the high variance of both utterance classes, i.e., genuine and spoofing classes. The proposed spoofing countermeasure addresses the poor generalization problem by identifying the hidden structure of each utterance category prior to the classification task. Specifically, fuzzy clustering algorithms were deployed to mine the hidden partitions of utterance classes. The conducted experiments showed that the proposed approach outperforms the state-of-the-art approaches in the ASVspoof 2017 dataset, with a testing EER equal to 1.07%.