Extreme Learning Machine-Based Unified Voice Spoofing Detection System
Shivam Devasar, Sanya Kukreja, Harsh Mishra, Vinal Patel · 2024
Spoofing attacks like the replay (i.e., using recorded voice of bonafide speaker), and cloning (i.e., using state-of-the-art Text-to-Speech for cloning bonafide voice and using) can be applied on various voice-assistant-enabled systems, voice-based biometrics, etc. Most of the existing countermeasures perform well on only specific attacks and thus fail to generalize for other classes of spoofing attacks. In this paper, an audio spoofing architecture using extreme learning machine (ELM) techniques is developed. Features such as acoustic local ternary patterns and gamma tone cepstral coefficients (GTCC) are employed. The performance of ELM is compared with Gaussian Naïve Bayes, decision tree classifier, and support vector machine (SVM) on the ASVspoof2019 dataset. It has been demonstrated in the result section that the proposed ELM-based spoofing detection system outperformed over mentioned schemes and achieved the highest accuracy of 99.2% with the lowest equal error rate value of 0.2% at a lower computation time.