The article discusses the problem of ‘audio deepfakes’, artificially produced or altered voice recordings that can convincingly imitate any individual’s voice. The authors present a new tool, the Multi-Feature Audio Authenticity Network (MFAAN), designed to detect such fabricated audio content. MFAAN incorporates multiple parallel paths to harness different audio representations and differentiate between genuine and manipulated recordings. Preliminary evaluations of MFAAN show its superior performance, with accuracies of 98.93% and 94.47% on two benchmark datasets.
Publication date: 8 Nov 2023
Project Page: not provided
Paper: https://arxiv.org/pdf/2311.03509
