Voice Liveness Detection KYC Project: Distinguishing Genuine and Spoofed Voices Using Deep Learning
Bugra Eyidogan, Gökberk Özsoy, Pedram Khatamino, Bilal Avvad, E. Judith Sen, Deniz Kumlu · Procedia Computer Science · 2025
Voice-based authentication systems are increasingly integral to secure user identification in various applications, from banking to smart devices. As these systems become more prevalent, their vulnerability to spoofing attacks, such as replayed or synthetic voices, poses significant security concerns. This paper addresses these challenges by focusing on speech recognition and voice liveness detection using advanced deep learning models. We explore the effectiveness of detecting pop noise—a low-frequency artifact characteristic of live speech—using the Constant-Q Transform (CQT) for feature extraction. Through a comprehensive analysis of datasets, including ASVspoof2015, ASVspoof2017, and POCO, we developed and trained a Convolutional Neural Network (CNN) model on the MeluXina supercomputer, achieving a test accuracy of 96.95% on the ASVspoof2017 dataset. Our findings demonstrate the potential of using pop noise detection and CNNs for robust voice liveness detection. Future work will expand on these results by exploring additional feature extraction methods and alternative machine learning algorithms to further enhance system reliability across various spoofing scenarios.