Replay spoofing attack detection using deep neural networks

In recent years, there has been increased interest in speaker verification(SV) systems and their usage has become widespread. This situation made the detecting of spoofing attacks, the discrimination of genuine speech from spoofed speech, an important research area for speaker verification (SV) systems. In this study, detection of replay spoofing attacks where a prerecorded speech signal is used to gain unauthorized access to ASV systems is studied. Mel frequency cepstral coefficients (MFCC) and long-term average spectrum (LTAS) statistics features are used to detect replay attacks using deep neural network (DNN) classifier. Experimental results using ASVspoof 2017 database show that MFCC and LTAS features with DNN classifier outperforms the Gaussian mixture model (GMM) classifier with constant Q transform cepstral coefficients (CQCC) which is the baseline replay attack detection system of the ASVspoof 2017 challenge.