Relevance of Quadrature Phase For Replay Detection in Voice Assistants (VAs)

There have been various studies involving Instantaneous Frequency (IF) estimation for Spoofed Speech Detection (SSD) task, such as the derivative of the phase obtained by Hilbert Transform (HT) approach and Energy Separation-based method. However, IF estimation by HT leads to lack of good temporal resolution. On the other hand, ESA-based method leads to excellent time resolution, however, lacks the relative phase information. Therefore, in this paper, we have proposed Cochlear Filter Cepstral Coefficients-based Instantaneous Frequency using Quadrature Energy Separation Algorithm (CFCCIF-QESA) feature set, which merits of having an excellent time resolution as well as inclusion of the relative phase information. Hence, we illustrate the significance of incorporating the quadrature-phase component along with the in-phase component for SSD of replay detection in VAs. To that effect, we perform experiments on the Realistic Replay AttackF Microphone-Array Speech Corpus (ReMASC) dataset. Furthermore, the proposed CFCCIF-QESA feature set gives 28.71 and 29.89 %EER using GMM and CNN respectively, on Eval set. The proposed feature set is evaluated using various performance metrics, including EER and other confusion matrix-based metrics. Finally, the latency for CFCCIF-QESA and CFCCIF-ESA is presented, showing the better suitability of the proposed CFCCIF-QESA feature set w.r.t. practical deployment.