Voice Liveness Detection Using Modified Group Delay Functions

Abstract

Automatic Speaker Verification (ASV) system is a key element of the voice biometric system. With the recent advancement in technology, spoofing attacks have also progressed drastically. In spoofing attack, attacker tries to get access of the biometric system by impersonating speech of genuine speaker. A successful attack would lead to serious consequences, such as stealing of information, fabricating false credentials for use in further attacks, gaining unauthorized access. In response to these attacks, researchers have come up with a solution to build more advanced ASV methods or develop effective spoofing countermeasures. The efforts has been made to develop algorithm to detect spoofing attacks effectively. Recently, new method has been developed to detect spoofing attacks known as Voice Liveness Detection (VLD). VLD has emerged as a successful technique to detect spoofing attacks in ASV system. VLD system verifies whether the input speech signal is from live speaker or it is the recorded samples by detecting popnoise in the speech. Pop noise comes out as a burst and is captured by the microphone as a breathing sound, which gets poorly produced by the loudspeakers. This noise acts as a cue for liveness in the input signal of an authentication system. VLD uses this strategy to distinguish between live or played speech. In this thesis, significance of the phase information in the signal is explored. The dynamics of the phase variation in the speech signal are represented using the group delay function. The feature sets developed using group delay function are used for VLD. The group delay function possess limitations for nonminimum phase signals and those limitations are overcome by Modified Group Delay (MGD) function. Two different approaches to develop modified group delay function are proposed, namely, Spectral Root (SR) and Linear Prediction (LP) smoothing-based group delay function. Performance of the proposed methods are evaluated on the POCO (POp noise COrpus) dataset. Moreover, Spectral Root Cepstral Coefficients (SRCC) has been also explored for the VLD task. In addition, the thesis work also proposes the Voice Privacy system-based on signal processing techniques as a countermeasure to prevent ASV system from various attacks.