Researchers propose a new method for verifying the source of deepfakes in speech. They claim that current methods assume speaker traits don't affect the verification process, but their study shows this isn't always true. To address this issue, they developed a framework called SDML (speaker-disentangled metric learning) which uses two novel loss functions to disentangle speaker and source information. The first function leverages Chebyshev polynomials to stabilize gradient optimization, while the second projects embeddings into hyperbolic space using Riemannian metric distances. Experimental results on a benchmark dataset show that SDML outperforms existing methods in certain scenarios. --- Why it matters: This matters because current deepfake detection methods may be flawed if they don't account for speaker traits. The proposed method could improve the accuracy of source verification, which is crucial for applications like voice assistant security and audio forensics. Source: https://arxiv.org/abs/2603.21875