Automatic Transcription of Organ Tablature Music Notation with Deep Neural Networks

Abstract

Organ tablature music notation differs considerably in structure and form from the music notation used today. The manual transcription of organ tablature compositions to modern music notation is time-consuming and often prone to errors. In this paper, we present a deep learning approach to automatically recognize organ tablature notation in scanned documents and transcribe it to modern music notation. Our approach is aimed at generating a uniform transcription that remains as close as possible to the original sheet music and therefore does not perform automatic error correction or musical interpretation. The artificial neural network model developed for the recognition of tablature characters is trained using a combination of real annotated tablature staves and tablatures produced by a synthetic data generator. The results of our experiments are evaluated on tablatures taken from two tablature books. We identify several types of error and validate that these are primarily caused by the poor legibility of relevant parts of some tablature scans. Overall, our approach achieves an accuracy of 97.2% and 99.3% correctly recognized bars, depending on whether note pitch and rest characters or note duration and special characters are considered, respectively.