Music structure analysis (MSA) aims to identify functional labels and their boundaries in music, a task that remains difficult despite recent progress. The paper from arXiv highlights three lingering challenges, starting with the scarcity of fine-grained annotations for training and evaluation. Existing systems also struggle with certain aspects of structure, though the abstract is cut off before detailing those remaining issues.

To address the first challenge, the authors introduce the SheetSage-MSA dataset, which provides a new source of fine-grained annotations. The paper's approach leverages both audio and lyric evidence, suggesting that combining these modalities can improve the detection of structural elements like verses, choruses, and transitions.

The dataset and the proposed use of lyric evidence are the core contributions. As the abstract is truncated, the full set of challenges and the exact methodology are not fully described here, but the focus on fine-grained data and multimodal evidence is clear from the title and abstract.