Hybrid AI Approaches for Improving 4-part Harmony Learning. Applications to Handwritten Optical Music Recognition.
| Tipo de publicação: | Inproceedings |
| Citação: | |
| Booktitle: | AI-LEAP 2026 - 1st International Workshop on AI Learning, Education And Practice, co-located with AIxIA 2026, University of Perugia, October 2026 |
| Ano: | 2026 |
| Mês: | October |
| Location: | Perugia |
| Resumo: | This paper presents a hybrid AI approach to handwritten Optical Music Recognition (OMR), aimed at supporting music students in learning four-part harmony. In previous work, we introduced Sharpmony, an AI-based mobile and web application adopted by more than 5,000 users, which has automatically checked over 21,000 harmony exercises against a set of classical harmony rules since 2020. To have an exercise checked, however, students must first create it in a music notation editor, even though harmony is traditionally written by hand in a music notebook. This step is both a barrier to adoption and a departure from the pen-and-paper practice that teachers value. Our goal is to remove it: students will thus write their exercises by hand, and a single photo is enough to produce the digital score that Sharpmony then checks. We present a pipeline that reconstructs the pre-printed staff from a photograph, locates each note head within the detected symbol, and reads its pitch from the fitted geometry, exporting a MusicXML score ready for automatic checking. On held-out pages it recovers 90.3% of the notes with both type and pitch correct; an end-to-end detector trained to predict the two jointly reaches 87.6%. No part of that reading is learned from data. Reading a score, in this setting, does not require learning to read: it requires detecting lines of the staff. The advantage of stating that knowledge explicitly is that it declares its own limits. Beyond the last printed line, the student draws the ledger lines by hand, and there the geometry no longer applies, but that boundary is itself measurable from the staff, so the notes at risk are known before any pitch is assigned and can be flagged for confirmation rather than misreported. By keeping practice on paper while preserving automated, immediate feedback, the approach lowers the technological barrier to AI-supported harmony learning in the conservatory and at home. |
| Palavras-chave: | Hybrid AI, Music Information Retrieval, Music Teaching Support, Optical Music Recognition, YOLO |
| Autores | |
| Adicionado por: | [] |
| Total mark: | 0 |
|
Anexos
|
|
|
Notas
|
|
|
|
|
|
Tópicos
|
|
|
|
|