Unseen Codec Spoof Speech Detection Based on Channel-Robust Feature

Abstract

For speech anti-spoofing, the ability of countermeasures (CMs) to cope with unseen attacks has been under scrutiny. Since the previous LA attack was mainly for ASV, which required that the spoofed speech be clean enough to be parsed properly by the ASV and that the unseen scenario be limited to the types of synthesis algorithms. With the development of DeepFake, spoofed speech is more often used to spread fake information so that the unseen codecs channel effects needs to be considered. Based on this, we propose a channel-robust spoof detection method based on the wav2vec2.0 and a channel augmentation adversarial (AUG-ADV) strategy. Our method was experimented on the FMFCC-A dataset and achieves the best results with several evaluation metrics.