Densely Connected Convolutional Network for Audio Spoofing Detection

摘要

Anti-spoofing has attracted increasing attention since the inauguration of the ASVspoof Challenges, due to the fact that automatic speaker verification (ASV) systems are vulnerable to spoofing attacks. The latest ASVspoof 2019 Challenge was dedicated to addressing attacks in three major classes: speech synthesis, voice conversion, and replay audio. In this paper, we propose a novel method that includes feature extraction, a densely connected convolutional network, and fusion strategies to answer the ASVspoof 2019 Challenge and to defend against spoofing attacks. Features are extracted using different algorithms and then fed separately into variants of our model, which differ only in terms of the kernel size of the global average pooling layer. A dense connectivity pattern with better parameter efficiency is introduced to the proposed network to strengthen the propagation of the audio features. The experimental results show that the proposed method improves the tandem decision cost function and equal error rate scores by 75% and 78%, respectively, in the logical access challenge. In the physical access challenge, the proposed method improves the t-DCF and EER scores by 73% and 72%, respectively, compared with state-of-the -art methods.