Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

MuseNet

openai.com

Post reply on HN
Print view View original

21–30 of 189 posts

Re: MuseNet

#21

Post by 0815test »

Not very impressive/interesting, they aren't releasing their code or models, and it seems that they had to use lots of rather obscure hacks to make the whole multi-instrument thing work. (Though I suppose it definitely is impressive in other ways, especially hacking their existing GPT architecture to do something worthwhile with non-text tokens. So, I'm not saying that any sort of work should be dismissed outright!)

I do still think that the good old https://github.com/hexahedria/biaxial-rnn-music-composition (Hexahedria's Biaxial RNN/Tied Parallel Networks, also published a paper achieving SOTA at least for that time, on a variety of midi datasets) has more interesting/compelling output musically, despite starting from a comparatively tiny dataset and using rather elegant and easily-undestandable convolution- and RNN-based techniques. Too bad that the implementation relies on Theano which is quite endangered at this point (doesn't seem to support up-to-date python 3?), but I do think it's a compelling starting point if anyone really wants to work on this domain.

Re: MuseNet

#22

Post by gambler »

rco8786 wrote: post #12

This seems incredibly applicable to musical scores in movies. I can imagine a product where the editor/director/someone inputs a handful of variables (mood, genre, instruments, etc) and timing requests (crescendo beginning at 30s and ending at 75s, calm period from 90s - 120s, etc) and out comes a musical score for the movie that matches up with their scene editing.

With hundreds of thousands of semi-professional and amateur musicians competing on the Internet for their 15 minutes of fame, getting film or computer game score just isn't a real problem.

Besides, this is already doable, without neural networks. E.g. Karma[1]. The fact that most "AI music" enthusiasts don't know/care about such systems is a clear indicator that they aren't really interested in music or helping creators and just want to shove AI in yet another niche where it doesn't belong, then pat themselves on the back for contributing to "progress".

Further commodification of what's left of out popular culture isn't something anyone should be excited about. I just can't wait for the endless wave of trite "style blends" and the inevitable "oh, but all human music is garbage anyway" justification.

[1] https://www.karma-lab.com/

Edit:

Also, some people don't seem to be aware of it either, but there is already a full-fledged genre of generative music where artists use anything from analog circuits[2] to randomized or algorithmic sequencers/arpeggiators[3] to custom-built digital devices[4] for creating entire tracks. Of course, the point of generative music usually isn't to replace the artist, it's to shift their focus.

[2] https://www.youtube.com/watch?v=u8Hr4wBRbaY

[3] https://www.youtube.com/watch?v=7n6-4buAkts

[4] https://www.youtube.com/watch?v=0KdTIUlGULk

Re: MuseNet

#24

Post by bduerst »

CamelCaseName wrote: post #17
craze3 wrote: post #5

Damn, I was just about to code this! Generative music is definitely on the come-up. If you like this, also check out https://generative.fm/ , which is from another HN member.

I think https://brain.fm also came out of HN. I've used it once or twice, but for whatever reason, nothing sounds better to me than the music I used to listen to when I was a teenager.

>nothing sounds better to me than the music I used to listen to when I was a teenager.

Tangental, but listening to the music from your teen years is a form of therapy for people who developed dementia. It's possible the music we impress in our teen years hold a special value in our brains.

Re: MuseNet

#26

Post by Areading314 »

This may be academically interesting, but the music still sounds fake enough to be unpleasant (i.e. there's no way I'd spend any time listening to this voluntarily).

Re: MuseNet

#27

Post by bgeeek »

rco8786 wrote: post #12

This seems incredibly applicable to musical scores in movies. I can imagine a product where the editor/director/someone inputs a handful of variables (mood, genre, instruments, etc) and timing requests (crescendo beginning at 30s and ending at 75s, calm period from 90s - 120s, etc) and out comes a musical score for the movie that matches up with their scene editing.

YouTube maybe, but movies? I could think of nothing worse. The score for a movie is just as important to me as the visuals.

Re: MuseNet

#28

Post by GistNoesis »

Nice.

We played a little with transformers inside a browser using tensorflow.js a few month ago, for real-time music transcription.

For those interested : Website : https://gistnoesis.github.io/ Project : https://github.com/GistNoesis/Wisteria/

My project is currently on hold, but will definitely receive update in the future. As it's kind of project for fun with no hope of monetization given that the space is already crowded with Google's Magenta, and now OpenAI is joining the dance with museNet.

Re: MuseNet

#29

Post by spaceheretostay »

Areading314 wrote: post #26

This may be academically interesting, but the music still sounds fake enough to be unpleasant (i.e. there's no way I'd spend any time listening to this voluntarily).

How does it sound fake? It sounds lovely to me, especially much of the Chopin+pop piece sampled in the article.

Re: MuseNet

#30

Post by dom96 »

Pretty cool. Ludum Dare is coming up and this would actually be a pretty nice way to generate some music.
Post reply on HN