Beyond Artifacts: Towards Generalizable Synthetic Song Detection via Music-Intrinsic Features

A new detection framework called Sofia can identify AI-generated songs with significantly higher accuracy than existing methods, according to research submitted to the arXiv preprint repository on June 15, 2026 [1][2]. The system improved the F1 score by 18.5 points over the strongest baseline on a newly constructed benchmark [2]. The framework, formally named “Synthetic-song detection framework via music features,” moves beyond conventional detection methods that rely on low-level audio artifacts or fixed assumptions about how generators work [2]. Instead, Sofia models music-intrinsic attributes — vocal characteristics, audio effects, and global structure — using feature-specific experts and an adaptive Mixture-of-Experts module [1][2]. The researchers configured Sofia with these representative feature sets and their combinations to measure their individual and complementary contributions [2]. To test the system, the team built MUSIC8K, a benchmark that incorporates the latest emerging AI music generators and realistic audio perturbations [1][2]. On the MUSIC8K-O subset, Sofia achieved the 18.5-point F1 improvement while maintaining robustness across different generator types [2]. The paper argues that this generator-agnostic representation is critical as AI music tools proliferate [2]. The work appears on arXiv, the open-access repository for electronic preprints that has hosted scientific papers since August 1991 [6]. arXiv now receives approximately 24,000 submissions per month and has surpassed two million total articles [6]. Papers on the platform are moderated but not peer-reviewed before posting [6]. The Sofia paper was submitted under the Computer Science > Sound category [1]. arXiv also supports community-built tools through its arXivLabs framework, which allows collaborators to develop experimental features such as citation explorers and code finders that appear on article pages [4][5]. The Sofia paper’s abstract page includes links to several of these tools, including Bibliographic Explorer and Connected Papers [1].