Replay attacks present a significant threat to Automatic Speaker Verification systems as they can be easily mounted using everyday smart devices by any non-professional imposter. The ASVspoof 2017 challenge was an initiative to develop solutions to counteract such replay attacks. The proposed solution builds on the fact that all the distinguishing features between genuine and spoofed audio are not effectively captured by conventional feature extraction techniques. Hence we propose a 1D ConvNet system with raw audio waves as features to it. This approach is able to achieve an EER of 0.41% on development set and 5.29% on evaluation set and hence outperforming best submission to ASVspoof 2017 challenge which had EER of 3.95% and 6.73% on development and evaluation sets respectively.
Paper
Full text
Replay attack detection with raw audio waves and deep learning framework
Semantic Scholar · Computer Science · 2019
Abstract
Replay attacks present a significant threat to Automatic Speaker Verification systems as they can be easily mounted using everyday smart devices by any non-professional imposter. The ASVspoof 2017 challenge was an initiative to develop solutions to counteract such replay attacks. The proposed solution builds on the fact that all the distinguishing features between genuine and spoofed audio are not effectively captured by conventional feature extraction techniques. Hence we propose a 1D ConvNet system with raw audio waves as features to it. This approach is able to achieve an EER of 0.41% on development set and 5.29% on evaluation set and hence outperforming best submission to ASVspoof 2017 challenge which had EER of 3.95% and 6.73% on development and evaluation sets respectively.