DEVELOPMENT OF A CLASSIFIER FOR OPTICAL RECOGNITION OF MUSICAL NOTATION

Gubenko Nadezhda, Molodyakov Sergey, Kolikova Tatiana

The representations that are used for two tasks related to sheet music are considered: identification of scores (musical notes) and obtaining the corresponding spectrograms and audio performances, taking into account the score as a search query. The scheme of the classifier is presented, the distinctive feature of which is the sequential use of several neural networks. The classifier is trained on a new data set of musical scores, which are collected from the DeepScores data set for segmentation, detection and classification of tiny objects, from the HOMUS data set - a pen-based musical notation and a data set of labeled GUITARPRO songs for sequence models. The combined data will become publicly available in the future. The results of extraction experiments using scans of real notes of high complexity are presented.

Keywords:recognition, musical notation, dataset, classifier, deep learning, spectrogram