SHNU Anti-spoofing Systems for ASVspoof 2019 Challenge

Abstract

This paper presents an experimental analysis of SHNU anti-spoofing systems for the ASVspoof 2019 challenge. This challenge focused on countermeasures for three major attack types, namely those stemming from the advanced technology of TTS, VC and replay spoofing attacks. According to the type of attacks, the challenge was divided into two independent sub-challenges, the logical access (LA) and physical access (PA). Results of different anti-spoofing technologies on both sub-challenges were reported. Furthermore, the same countermeasures were also evaluated on two previous challenges, the ASVspoof 2015 and 2017. Experiments on cross-databases showed that, it appeared hard to generalize the classifiers trained from ASVspoof 2019 LA and PA databases to the previous challenges. The generalization ability of anti-spoofing technologies to different, new and unknown conditions was still very challenging. In addition, the effectiveness of different acoustic features were also examined and reported. Finally, we investigated the linear and an interfusing score-level fusion methods to individual systems to achieve better performance.