Morse Wavelet Features for Pop Noise Detection

Abstract

Spoofed Speech Detection (SSD) problem has been an important problem, especially for Automatic Speaker Verification (ASV) systems. However, the techniques used for designing countermeasure systems for SSD task are attack-specific, and therefore the solutions are far from a generalized SSD system, which can detect any type of spoofed speech. On the other hand, Voice Liveness Detection (VLD) systems rely on the characteristics of live speech (i.e., pop noise) to detect whether an utterance is live or not. Given that the attacker has the freedom to mount any type of attack, VLD systems play a crucial role in defending against spoofing attacks, irrespective of the type of spoof used by the attacker. To that effect, we propose Generalized Morse Wavelet (GMW)-based features for VLD, with Convolutional Neural Network (CNN) as the classifier at the back-end. In this context, we use pop noise as a discriminative acoustic cue to detect live speech. Pop noise is present in live speech signals at low frequencies (typically $\leq 40$ Hz), caused by human breath reaching at the closely-placed microphone. We show that for $\gamma =3$, the Morse wavelet has the highest concentration of information denoted by the least area of the Heisenberg’s box. Hence, we take $\gamma =3$ for our experiments on Morse wavelets. We compare the performance of our system with Short-Time Fourier Transform (STFT)-Support Vector Machine (SVM)-based original baseline, and other existing systems, such as Constant Q-Transform (CQT)-SVM, STFT-CNN, and bump wavelet-CNN. With overall accuracy of 86.90% on evaluation set, our proposed system significantly outperforms STFT-SVM-based original baseline, CQT-SVM, STFT-CNN, and bump wavelet-CNN by an absolute margin of 18.97 %, 8. 02%, 15. 09%, and 12. 21%, respectively. Finally, we have also analyzed the effect of various phoneme types on VLD system performance.