One-class network leveraging spectro-temporal features for generalized synthetic speech detection

Jiahong Ye, Diqun Yan, Songyin Fu, Bin Ma, Zhihua Xia

Abstract: Highlights•We designed a spectro-temporal network to capture F0 subband differences.•We improved OC-Softmax with the KoLeo regularizer to enhance intra-class balance.•Experiments on ASVspoof 2019 LA and 2021 LA show better results of our method.