Voice Liveness Detection Using Linear Frequency Residual Cepstral Coefficients

This study deals with the problem of voice liveness detection (VLD) in order to increase biometric security. We propose Linear Frequency Residual Cepstral Coefficients (LFRCC) features, and Bi-directional Long Short-Term Memory (BiLSTM) classifier to ascertain whether the speech is live or not. We employed the POp COrpus (POCO) dataset, which is a standard dataset for the VLD problem. We compare our work with various other well known spectral features, such as Mel Frequency Cepstral Coefficients (MFCC), and Linear Frequency Cepstral Coefficients (LFCC) using BiLSTM as pattern classifier. Our optimal results indicate 83.14 % accuracy; which is 20.85 % greater than the existing STFT-based approach. To investigate practical suitability of VLD system using the proposed LFRCC feature set, analysis of latency period, feature-fusion, and comparison with pre-trained models is also presented in this study.