OpenAI recently released access to a music generation tool based on MuseNet. MuseNet is an unsupervised deep neural network trained on “hundreds of thousands of MIDI files”, and just like the way GPT-2 was trained to generate text, MuseNet can generate music.
I wanted to check out what the tool could do—this is a song it composed from no intro notes, in the style of video game music using all available instruments and maximum tokens per generation step:
Pretty good, right? This would very nearly pass for an actual game soundtrack. There are a few oddities, especially the seeming randomness of note and section lengths. There’s perhaps a little less musical direction than a human-composed song, since it doesn’t really seem to be heading anywhere specific…but I could be imagining things.
For comparison, here’s a composition I made based on what MuseNet produced. The first few notes are taken right from its composition, but then I went off in a slightly different direction. The main thing I did was clean up the rhythmic changeovers—MuseNet tended to seem a little random in introducing new sections or instruments.
I’m curious to see how these kinds of music-generation tools will affect the landscape of working composers. I definitely wouldn’t have written a song exactly like this if I started from scratch, but I did add a little creative flair to the piece. I’m not nearly as talented as an actual composer, so having this kind of heavy “inspiration” is helpful for me.
Maybe the way tools like this will be used is as little snippets of inspiration, or things to be sampled—not entirely unlike the way musicians are able to take bits and pieces of other human-composed songs now.
Then again, maybe I just got lucky and happened to hit on a good tune. This was the second one I generated from scratch, with the same settings:
Not so great. There’s some material that could be transformed into a really catchy song, but it’d definitely need some cleaning up. For now, then, this tool is great for inspiration, but isn’t gonna be popping out one-hit wonders.