New AI retrieves music scores by content, not just metadata
Transformers and LLMs now search sheet music by melody and notes
A team led by Noelia Luna-Barahona (University of Alicante) has published a paper titled "Direct content-based retrieval from music scores images" on arXiv (arXiv:2605.22255). The work addresses a critical gap in musical information retrieval: while text documents have long benefited from content-based search, music scores still rely primarily on metadata (title, composer). The researchers first study which features of a score are most relevant for search and define a systematic method to build query datasets from any annotated corpus.
They then compare three approaches: traditional Optical Music Recognition (OMR) pipelines that transcribe the score before searching, a transcription-free Transformer model trained to recognize queries directly from score images, and a text-prompted Large Language Model. Evaluated on four diverse corpora (varying in size, image quality, and typesetting), they find that OMR-based methods achieve higher accuracy when the target domain matches training data, whereas transcription-free models are more robust across different domains. The work opens the door to practical content-based retrieval tools for musicians, educators, and researchers—letting them search by musical phrase, rhythm pattern, or note sequence without manual transcription.
- OMR-based pipelines achieve higher in-domain retrieval accuracy but struggle with domain shifts
- Transcription-free Transformer model handles domain variability better, offering more robust cross-corpus performance
- Researchers introduce a systematic method to create query datasets from any annotated music score corpus
Why It Matters
Enables musicians and scholars to search scores by melody, rhythm, or notes—not just title or composer.