Attention-Based Convolutional Neural Network for ASV Spoofing Detection

In recent years, automatic speaker verification (ASV) algorithms have undergone significant progress. They have been widely deployed in different applications, but the ASV systems are vulnerable to spoofing attacks, such as impersonation, replay, text-to-speech, voice conversion and the recently emerged adversarial attacks. To improve the robustness of the ASV system, researchers have designed anti-spoofing systems to resist spoofing attacks. While previously proposed systems have shown to be effective for spoof attacks detection, they are all ensemble methods based on different speech representations and architectures at the cost of increased model complexity, with similar performance not being achieved with single systems. This paper proposes an attention-based single convolutional neural network to learn discriminative feature embedding for spoof detection, achieving performance comparable to ensemble methods. The key idea is to decrease the information redundancy among channels and focus on the most informative sub-bands of speech representations. The experiments show that our proposed single system achieves an equal error rate of 1.87% on the evaluation set of ASVspoof 2019 Challenge, outperforming all single systems and comparable to the second-ranked system (EER 1.86%) among all known systems.