Non-Cepstral Uncertainty Vector for Replay Spoofed Speech Detection

Replay spoofing attack refers to gaining unautho-rized access to Automatic Speaker Verification (ASV) system simply by replaying the pre-recorded speech. In order to make an ASV system immune towards such attacks, this paper exploits Heisenberg's uncertainty principle in signal processing frame-work for spoof speech detection (SSD). This principle is used to develop non-cepstral features, namely, u-vector, t-vector, and
$\omega$
-vector for SSD task. On analyzing the pictorial representation of the proposed feature sets, it is concluded that t-vector is effective at detecting distortions due to replay channel and reverberation. On the other hand,
$\omega$
-vector is although less effective for SSD but possess an inherent ability to emphasize more on the high frequency regions than the low frequency regions, which is desirable for effective SSD. Their summation, u-vector adopts properties of both t-vector and
$\omega$
-vector and thus, proving it to be superior for SSD task. Another attribute of these feature representation lies in their simplicity of implementation for practical SSD deployment. Experiments performed on ASVspoof 2017 version 2.0 dataset using Gaussian Mixture Model (GMM) classifier suggest that u-vector gives an absolute reduction of 4.55% and 5.28% in EER on development and evaluation dataset, respectively, on the baseline system. Automatic Speaker Verification (ASV), Heisenberg's un-certainty principle, Time-Bandwidth Product (TBP), replay attack, u-vector.