Voice liveness detection under feature fusion and cross-environment scenario

摘要

Detecting playback spoofing attacks in speaker verification system is a big challenge. Recent studies on ASVspoof challenges show that replay attacks are the most difficult to recognize. Reasonable performance is expected from such antispoofing systems to avoid malicious access attempts on voice biometrics enabled systems for possible commercial deployment. We present a study on filterbank based short-term cepstral features for liveness detection to counter replay spoofing attacks on speaker verification systems. These systems are evaluated on ASVspoof 2017 version 2.0 dataset. Experimental investigation is carried out on standalone and fused features to assess the performance of the antispoofing systems using spoofing detection equal error rate (EER). Improvement of 20.47% and 21.51% is obtained over baseline system using standalone and fused approaches, respectively. We also explore the impact of proposed static inverted Mel frequency cepstral coefficients (1MECC) based system under mismatched conditions by training and testing it in different environments (with different background conditions) alongwith other systems. Results show that the proposed system outperforms other systems used in this study in all experiments.