The recognition of patterns that have a time dependency is common in areas like speech recognition or natural language processing. The equivalent situation in image analysis is present in tasks like text or video recognition. Recently, Convolutional Recurrent Neural Networks (CRNN) have been broadly applied to solve these tasks in an end-to-end fashion with successful performance. However, its application to Optical Music Recognition (OMR) is not so straightforward due to the presence of different elements sharing the same horizontal position, disrupting the linear flow of the timeline. In this paper, we study the ability of the state-of-the-art CRNN approach to learn codes that represent this disruption in homophonic scores. In our experiments, we study the lower bounds in the recognition task of real scores when the models are trained with synthetic data. Two relevant conclusions are drawn: (1) Our serialized ways of encoding the music content are appropriate for CRNN-based OMR; (2) the learning process is possible with synthetic data, but there exists a glass ceiling when recognizing real sheet music.

State-of-the-art Optical Music Recognition (OMR) techniques follow an end-to-end or holistic approach, i.e., a sole stage for completely processing a single-staff section image and for retrieving the symbols that appear therein. Such recognition systems are characterized by not requiring an exact alignment between each staff and their corresponding labels, hence facilitating the creation and retrieval of labeled corpora. Most commonly, these approaches consider an agnostic music representation, which characterizes music symbols by their shape and height (vertical position in the staff). However, this double nature is ignored since, in the learning process, these two features are treated as a single symbol. This work aims to exploit this trademark that differentiates music notation from other similar domains, such as text, by introducing a novel end-to-end approach to solve the OMR task at a staff-line level. We consider two Convolutional Recurrent Neural Network (CRNN) schemes trained to simultaneously extract the shape and height information and to propose different policies for eventually merging them at the actual neural level. The results obtained for two corpora of monophonic early music manuscripts prove that our proposal significantly decreases the recognition error in figures ranging between 14.4% and 25.6% in the best-case scenarios when compared to the baseline considered.