Improved Processing of LP-residual Information for Detection of Replay Signals
Krishna Dutta, Madhusudan Singh, Debadatta Pati · 2019
Replay attack is a method of unauthorized access to automatic speaker verification system by using pre-recorded speech samples of any target. Recently, linear prediction (LP) residual signal is extensively explored for detection of replay signals. The latest development is the mel-cepstral processing of the analytic representation of the LP residual signal. In that work, the conventional cepstral processing method is followed and the effect of zeroth coefficient (c0) is ignored. In this work, the effect of c0is investigated. In particular, we have used residual Hilbert envelope mel frequency cepstral coefficient (RHMFCC) and residual phase-mel frequency cepstral coefficient (RPMFCC) feature to design the replay detection system. The experiments are carried out in ASVspoof 2017 version 1.0 database. With the incorporation of c0, the performance of RHM-FCC based system is degraded slightly (around 1%), but the performance of RPMFCC based system is improved significantly. This improvement is significant enough to provide a relative improvement of 11.78% in the case of combined RHMFCC+RPMFCC based system with c0compared to the system without c0. The system performance with constant-Q cepstral coefficient with log energy (CQCC-L) feature is improved slightly from 12.72% to 12.47% with c0. The final system of CQCC-L+RHMFCC+RPMFCC feature with c0shows a relative improvement of 20.51% compared to the system without c0. These results show the importance of c0in excitation signal based replay detection system.