Recent advances in Deep Learning have propelled the development of fields such as Optical Music Recognition (OMR), which is responsible for extracting the content from music score images. Despite progress in the field, existing literature scarcely addresses core issues like performance in real-world scenarios, user experience, maintainability of multiple pipelines, reusability of architectures and data, among others. These factors result in high costs for both users and developers of such systems. Furthermore, research has often been conducted under certain constraints, such as using a single musical texture or type of notation, which may not align with the end-user requirements of OMR systems. For the first time, our study involves a comprehensive and extensive experimental setup to explore new ideas towards the development of a universal OMR system---capable of transcribing all textures and notation types. Our investigation provides valuable insights into several aspects, such as the ability of a model to leverage knowledge from different domains despite significant differences in music notation types.