SNAP: Speaker Nulling for Artifact Projection in Speech Deepfake Detection

To address the issue of insufficient generalization in speech detectors due to the “speaker entanglement” phenomenon in self-supervised speech encoders, the research team proposed the SNAP (Speaker Nulling for Artifact Projection) framework. This framework effectively suppresses speaker-dependent components in features by estimating the speaker subspace and applying orthogonal projection techniques, allowing the detector to focus on cues related to synthetic artifacts. Experiments show that this method achieves advanced detection performance on unseen speakers, significantly improving the generalization ability of speech deepfake detection.