Optical Music Recognition with deep learning

Context

Music plays a important role in my family, we all practise music instrument and I learned how to read music long ago. During a lecture at TU Wien, I discovered the dr. Pacha was involved in a field of research combining the two world of data science and music. I directly though about doing my master thesis under he’s supervision. Also what make more sens than studying music in Vienna?

Project

Lot of music notation exist in a format computer can only display but not read. Optical Music Recognition (OMR) is a field of research, which goal is to turn non-computer readable score into computer readable scores. With these scores, we would be able to listen to the music, edit the scores, make queries (e.g. ask if a pattern is typical of a compositor)

A mean to understand music notation is to imagine some primitive element interact to form the notes. My goal was to use Graph Neural Networks (a type of machine learning model) to recover these interactions and build so-called Music Notation Graphs (see illustration).

This project was definitely challenging and I encourage you to read my master thesis for more details. It concluded my academic journey and also led me to my first paper as first author (link will be added soon).