Among us - Detecting spoofed audio samples in a multi-party conversation -
Author
Anandhamurugan, Vigneshwar
Term
4. semester
Education
Publication year
2025
Submitted on
2025-06-03
Pages
78
Abstract
As voice technologies become more prevalent, it becomes vital to improve spoofing detection algorithms especially when multiple speakers are involved. This thesis presents a novel pipeline combining speech separation with spoof detection to address the challenge of identifying spoofed audio samples in overlapped speech. The system uses a SOTA spoof detection model which achieves an EER of 2% on ASVspoof 2019 dataset. However, the EER drops to 24% when tested on a custom- generated mixed audio dataset, as a result of modifications of the audio artifacts by the speech separation process. The system when trained on the custom dataset saw an increase in EER to 22% highlighting the importance of domain adaptation for spoof detection in complex acoustic environments and provides a foundation for future research in real-world scenarios.
Documents