Algorithms and new frontiers in AI music

[Timzoid]

I often wonder what Mozart would think of the music of Debussy. He might not like it at first, but I'm sure he'd be fascinated with it. I think there is still that future in music, to create something new, that would please humans, but be refreshing and different. AI can help explore that new territory, expand music into different styles.

When I was working with MuseNet, I usually worked on at least three compositions at once, open in different tabs, so I was generating a lot of measures. Working in multiple tabs was just a way of not waiting. Sometimes it would generate a phrase or a few measures of music that were surprising, out of character for the style used, and something I thought few or no composer would think of writing, yet it somehow fit. I used one such phrase in my piece, The Code. It's not inhibited like humans are.

To me, the freshness of what MuseNet would come up with at times is comparable to the things that LLMs could do, that surprised even the people working on the LLMs. Anyway, in LAMC, I would like there to be a feature that turns up that factor, generating things that are unexpected, where perhaps part of the algorithm would generate something random, and another part would take the random and bring coherence to it. Since I don't know anything about algorithms, it's just an idea.

Well, it's only my second week of using LAMC, so it's going to take me a few months to feel more confident about settings and the type of seeds which produce the best music.

[asigalov61]

@Timzoid Yes, MuseNet was indeed very good, especially because of the interface and free compute resources (hosted by OpenAI). I also used MuseNet heavily as you can see on my SoundCloud. I also used it to compose several compositions simultaneously, so when it went down my music production sorta slowed down since I have to use my own models that are slower and require a GPU workstation for each instance.

As far as generating something unexpected...there are several ways to do it. You can either use inpainting of a particular instrument in LAMC or there are other projects (not mine) that can do partial inpainting of the music composition. Like this one by Stanford researchers.

https://github.com/jthickstun/anticipation

Otherwise, I completely agree with you about what you wrote. MuseNet will be missed by those who appreciated it.

Definitely try Euterpe (the large model in particular) and also Lars Ulrich Transformer (on your Piano pieces for example).

And thank you again for using LAMC. I really appreciate it and I am happy that it works for you.

Alex.

PS. I agree that OpenAI took down MuseNet probably because they either ran out of budget for it or because they wanted to allocate all resources to ChatGPT/DALL-E. I do not think it was a copyright issue since MuseNet was made with pub MIDIs and because MIDIs is not a DRM format (IMHO of'course).