DeepTab: Multi-Line Organ Tablature Transcription

This repository provides the trained models and the datasets presented in the paper "Automatic Transcription of Organ Tablature Music Notation with Deep Neural Networks (Schneider et al., 2021, http://doi.org/10.5334/tismir.77).

The data consists on the one hand of real tablature staves extracted from the scans of two organ tablature books:

• "Orgel oder Instrument Tabulaturbuch" (’Organ or InstrumentTablature Book’) from 1583
• "Ein new künstlich Tabulaturbuch" (’A new artificial tablaturebook’) from 1575

Both books were written by German Organist and Composer Elias Nikolaus Ammerbach. We manually annotated 1200 staves from each book with label sequences. To increase the amount of available data we employed data augmentation and a synthetic data generator that randomly arranges images of single characters into tablature rows. The whole data set requires almost 90GB of disc space, which is why instead of the whole data set we distribute our generation and augmentation tools along with the required source images. We also provide a script to download the annotated real tablature staves.

The best models from our papers are available in the Tensorflow2 SavedModel format (https://www.tensorflow.org/guide/saved_model). We provide a code snippet to load the models and perform predictions. We also provide a notebook that allows the transcription of whole pages of scanned organ tablatures.

File Structure

The repository is structured as follows:

• models contains download scripts for the best model from our paper and an improved attention model as well as a code snippet to perform predictions with these models.
• data contains download scripts for our real data sets as well as the source images used by our data generator.
• generator contains the dataset generator scripts allowing to create artificial organ tablature rows from a given set of source images.
• transcription contains scripts for transcribing whole scanned organ tablature pages.