The sheet music is input as an image, which is not the right encoding for comparison with MIDI, so you need some kind of image recognition in the pipeline. I do

The sheet music is input as an image, which is not the right encoding for comparison with MIDI, so you need some kind of image recognition in the pipeline. I do think that they could replace the recurrent part of their network by dynamic programming and only use machine learning for feature extraction.