End-to-end audio replay attack detection using deep convolutional networks with attention

Abstract

With automatic speaker verification (ASV) systems becoming increasingly popular, the development of robust countermeasures against spoofing is needed. Replay attacks pose a significant threat to the reliability of ASV systems because of the relative difficulty in detecting replayed speech and the ease with which such attacks can be mounted. In this paper, we propose an end-to-end deep learning framework for audio replay attack detection. Our proposed approach uses a novel visual attention mechanism on time-frequency representations of utterances based on group delay features, via deep residual learning (an adaptation of ResNet-18 architecture). Using a single model system, we achieve a perfect Equal Error Rate (EER) of 0% on both the development as well as the evaluation set of the ASVspoof 2017 dataset, against a previous best of 0.12% on the development set and 2.76% on the evaluation set reported in the literature. This highlights the efficacy of our feature representation and attention-based architecture in tackling the challenging task of audio replay attack detection.