Improved One-class Learning for Voice Spoofing Detection

Lixiang Li, Xiaopeng Xue, Haipeng Peng, Yeqing Ren, Mengmeng Zhao

Abstract: Most of the studies on one-class learning for voice spoofing detection have poor generalization ability for unknown attacks and lack flexibility for different tasks. To address these problems, we propose an improved one-class softmax (IOC-Softmax) by reformulating the scale factor of one-class softmax (OC-Softmax) into two parts such that it down-weights the loss assigned to bona fide samples. Our IOC-Softmax is a novel dynamically scaled OC-Softmax, which can select different scale factors for different tasks, so that it can prevent a large number of bona fide samples from overwhelming the anti-spoofing model during training. To verify the effectiveness of the proposed IOC-Softmax, we perform numerous experiments on the ASVspoof 2019 dataset. The results show that our IOC-Softmax achieves better performance when compared with OC-Softmax.