Anti-Spoofing Using Transfer Learning with Variational Information Bottleneck

Abstract

Recent advances in sophisticated synthetic speech generated from\ntext-to-speech (TTS) or voice conversion (VC) systems cause threats to the\nexisting automatic speaker verification (ASV) systems. Since such synthetic\nspeech is generated from diverse algorithms, generalization ability with using\nlimited training data is indispensable for a robust anti-spoofing system. In\nthis work, we propose a transfer learning scheme based on the wav2vec 2.0\npretrained model with variational information bottleneck (VIB) for speech\nanti-spoofing task. Evaluation on the ASVspoof 2019 logical access (LA)\ndatabase shows that our method improves the performance of distinguishing\nunseen spoofed and genuine speech, outperforming current state-of-the-art\nanti-spoofing systems. Furthermore, we show that the proposed system improves\nperformance in low-resource and cross-dataset settings of anti-spoofing task\nsignificantly, demonstrating that our system is also robust in terms of data\nsize and data distribution.\n