Beyond Artifacts: Towards Generalizable Synthetic Song Detection via Music-Intrinsic Features
Researchers have developed Sofia, a new framework for detecting AI-generated songs that moves beyond analyzing low-level artifacts. Sofia utilizes music-intrinsic attributes through a Mixture-of-Experts (MoE) module, incorporating features like vocal, audio effects, and global structure. To test its effectiveness, a new benchmark called MUSIC8K was created, featuring current AI music generators and realistic audio perturbations. Experiments demonstrated that Sofia achieves generator-agnostic representations, significantly improving detection accuracy and robustness. AI