Music Representation Learning for Efficient Multimodal Retrieval

Music archives are inherently multimodal: the same work may exist as audio, scanned sheet images, and symbolic scores. Yet most Music Information Retrieval (MIR) systems assume a single curated modality, limiting cross-modal search and real-world deployment. We address this gap by learning a unified embedding space that aligns music across modalities, enabling multimodal retrieval with compact, efficient representations. Experiments on an extended version of the MUSCUTS dataset show promising composer-level cross-modal retrieval, particularly for audio and symbolic queries.