Significance of Constant-Q Transform for Voice Liveness Detection
Kuldeep Khoria, Ankur T. Patil, Hemant A. Patil
Abstract
In this paper, we present the novel approach of the liveness detection in speech signal based on constant-Q transform (CQT) which employs geometrically distributed frequency bins. Pop noise can be attributed to the liveness in the speech signal and we exploited this attribute for liveness detection. Pop noise is created due to spontaneous breathing while uttering the certain phonemes which includes the plosive burst, and it has low frequency characteristics. We follow the approach of liveness detection in original POCO dataset paper as baseline, where features are derived from Short-Time Fourier Transform. In our approach, we exploited the low frequency characteristics of pop noise using CQT which has variable spectro-temporal resolution with high resolution at low frequency regions. The experiments are performed on recently released publicly available POp noise COrpus (POCO) dataset. The 10-fold cross-validation performed using proposed approach shows improvement in absolute accuracy by 4.2% as compared to the baseline system. The proposed approach also shows relatively better performance for our disjoint partition (in terms of speakers) of the dataset.