I have a soundtrack and each song has 16 channels/instruments. Nearly every song uses different instruments. For a project I'm working on, I need to have channels for every instrument used in the entire soundtrack, in every song. Does something like this already exist or do I have to create it?
The goal is to train a recurrent neural network on the FFVII soundtrack to generate more classic FFVII music. The state-of-the-art either trains on WAV files which requires more computing resources than I have, or MIDI but the corpus has to be restricted to a consistent set of instruments. If I transformed each MIDI file into a sparse binary 3D matrix where the three dimensions represent instrument type, time step, & note, I'd be able to get some decent results. Could I perhaps extract the instrument code for each track, & use that to build my 3D matrices? I can already extract notes & times.