Energy Separation Based Instantaneous Frequency Estimation from Quadrature and In-Phase Components for Replay Spoof Detection

For replay Spoof Speech Detection (SSD), features that incorporate auditory transform-based information as well as Instantaneous Frequency (IF) information have been proposed in the past. IF is estimated either by derivative of analytic phase via Hilbert transform, or by using high temporal resolution Teager Energy Operator (TEO)-based Energy Separation Algorithm (ESA). However, the excellent temporal resolution of ESA comes with lacking in using relative phase information, and vice-versa. Hence, we propose novel CFCCIF-QESA features, with excellent temporal resolution as well as relative phase information. CFCCIF-QESA is designed by exploiting relative phase shift, without estimating phase explicitly. Effectiveness of proposed approach is validated by mutual information and Kullback-Leibler (KL) divergence-based analysis. Furthermore, TEO is used for complex signals for SSD. Consequently, the novel ideas of quadrature relative phase and TEO for complex signals are exploited for improving the performance of CFCCIF-ESA on ASVspoof 2017 version2.0 and ASVspoof 2019 databases. On ASVspoof 2017 evaluation set, when compared with CQCC features, CFCCIF-QESA features yield percentage improvement of 35.51% and 30.19%, with GMM and CNN classifiers, respectively. As compared to CFCCIF-ESA, a percentage improvement of 30.40% is achieved on ASVspoof 2019 evaluation dataset with GMM. Finally, the analysis of latency period indicates relatively better performance of CFCCIF-QESA and thus, its potential for practical SSD system deployment.