A study on data augmentation in voice anti-spoofing

In this paper we perform an in depth study of how data augmentation techniques improve synthetic or spoofed audio detection. Specifically, propose methods to deal with channel variability, different compressions, bandwidths and unseen spoofing attacks. These challenges, have all been shown significantly degrade the performance based systems anti systems. Our results are on ASVspoof 2021 challenge, Logical Access (LA) Deep Fake (DF) categories. is Data-Centric, meaning that models fixed by manipulating data. We introduce two forms - compression for DF part, LA part. addition, a double sided log spectrogram feature design improves centering sub-bands interest, where discriminating artifacts can be localized. Furthermore, new type online augmentation, SpecAverage, introduced. This method includes masking features their average value order generalization. best single system fusion schemes both achieve state art category, EER 15.46% 14.27%, respectively. task reduced baseline 50% min t-DCF 16%. from wide variety distributions replicated help speech enhance results.