41–50 of 59 posts
https://www.musicdevelopments.com/ Technically not ‘AI’ though, more based on music theory ‘rules’.
Of course there exists rule-based AI. Whatever works to partially replace a human professional.
That said, I don't have anything to show yet... Only got a landing page at https://tunesage.com at the moment... but here's an ancient prototype writing very generic melodies just to prove I'm trying... https://www.youtube.com/watch?v=8cwEaiwEZfU
Regardless of my personal eventual success or failure in the endeavour, the field does seem to be ripe for innovation! Nothing out there at the moment (that I've seen at least) seems to satisfy my current desires as a composer.
I've been working with OpenAi's jukebox to make vocals / "singer-songwriter" type stuff rather than electronica. Can do a HN writeup on process if there is interest... Youtube Playlist: https://www.youtube.com/watch?v=_lcCJzfXl50&list=PLaa32nLgVv... SoundCloud: https://soundcloud.com/songshtr/sets/machinecroon Github: https://songshtr.github.io The main challenge for me so far has been post-processing. Jukebox produc…
I've been working with OpenAi's jukebox to make vocals / "singer-songwriter" type stuff rather than electronica. Can do a HN writeup on process if there is interest... Youtube Playlist: https://www.youtube.com/watch?v=_lcCJzfXl50&list=PLaa32nLgVv... SoundCloud: https://soundcloud.com/songshtr/sets/machinecroon Github: https://songshtr.github.io The main challenge for me so far has been post-processing. Jukebox produc…
That's impressive. And the music is much more different than I expected, in a good way. So you wrote the lyrics? Or is it autogenerated as well?
Yes I wrote the lyrics myself (other than a couple quotes from Eliot & Borges) - no GPT etc involved in the text. Even with the tunes, the seed music fed to Jukebox gives some control over what gets generated next... especially when you later splice various outputs together. Each ~2-3 minute song needs anywhere between 15-100 2-3 minute raw snippets to be generated from Jukebox before you can find enough sections to paste together. I suppose I would be considered a writer/producer while Jukebox is clearly the performer - an immensely talented, yet completely capricious one!
The thing that's missing is the data. If we had midi transcriptions of 100k songs (abc notation could be fine too), we could probably get really interesting stuff, but most of what is available is lossy chord transcriptions and classical music (public domain). So if you want to automatically create something that sounds like mozart, you're in luck!
But this isn't really satisfying to me. For generative music, we're still largely stuck with encoding musical rules in code rather than feeding data to a transformer. To me, the former feels much less like AI than the latter. The data is all locked up behind an impossible quagmire of copyright.
But if I were a sheet music publishing company, I would be seriously considering the future of music creation with AI given my broad access to notated music & metadata (is this an original score, or a grade 1 simplification?). But again, music copyright is a pretty complex contraption.
Has anyone solved the data problem? I have taken many stabs at this over the last 10 years, and the primary problem is that music is _extremely_ subtle. Even something as seemingly simple as chord progressions are very difficult to annotate properly without deep training. And even then, experts can disagree. The thing that's missing is the data. If we had midi transcriptions of 100k songs (abc notation could be fine…
So MIDI songs were being generated quite "early" on, even with basic text generators (LSTM models), where the "alphabet" was replaced by MIDI symbols. [1]
What Jukebox (and other models) did well was work with raw audio, rather than rely on MIDI or similar transcriptions. They break down audio into "blocks" using Fast Fourier Transforms, then each block is "tokenized" and then fed into transformers, similar to GPT-3 or other newer text generation models. This now allows the "subtle" musicalities to be discovered by models without relying on transcription. Wavenet (from Google) [2] I believe was the first one to do this... I tried myself for 6 months, but the processing power required for this is candidly only available for the Google/OpenAIs of the world.
>> the data is all locked up behind an impossible quagmire of copyright
So my understanding is that the training of Jukebox is similar to Copilot. It was trained on the entire English Spotify catalogue so all that has been digested by the model - under Fair Use. But again - this is probably far from "settled" legally.
[1] https://arxiv.org/abs/1612.07837
[2] https://www.deepmind.com/blog/wavenet-a-generative-model-for...
Earlier quoted context omitted.
That's impressive. And the music is much more different than I expected, in a good way. So you wrote the lyrics? Or is it autogenerated as well?
Thanks so much for your kind words... and for taking the time to listen! Yes I wrote the lyrics myself (other than a couple quotes from Eliot & Borges) - no GPT etc involved in the text. Even with the tunes, the seed music fed to Jukebox gives some control over what gets generated next... especially when you later splice various outputs together. Each ~2-3 minute song needs anywhere between 15-100 2-3 minute raw snip…