Voice Liveness Detection using Constant-Q Transform-Based Features

In this work, we propose to use the Constant-Q transform (CQT)-based feature set for voice liveness detection (VLD), which can enhance the confidence in authenticity of the speaker in Automatic Speaker Verification (ASV) system. The live speaker can be characterized via his/her voice using the presence of the pop noise in the speech signal. Pop noise comes out as a burst and possesses the low frequency characteristics. In this paper, we present the modified CQT-based approach over the traditional Short-Time Fourier Transform (STFT)-based algorithm (baseline) for VLD. The experiments are performed on recently released POp noise COrpus (POCO) dataset with various statistical, discriminative, and deep learning-based classifiers, namely, Gaussian Mixture Models (GMMs), Support Vector Machine (SVM), Convolutional Neural Networks (CNN), and Light-CNN (LCNN), respectively. The significant improvement in performance is observed for the proposed CQT-based features over STFT-based features. Relatively best performance is obtained for CQT-LCNN architecture, which shows 81.93% accuracy on evaluation set. Furthermore, we analyzed the performance of the CNN and LCNN-based VLD systems for each word using proposed CQT-based vs. STFT-based baseline features.