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

MusicLM: Generating music from text

arxiv.org

Post reply on HN
Print view View original

71–80 of 114 posts

Re: MusicLM: Generating music from text

#71

Post by dwringer »

netr0ute wrote: post #67
bulbosaur123 wrote: post #62

I'm absolutely flabbergasted by rapid progression of music related machine learning models. I've always seen music as the "final frontier". With images you can have tiny errors and noise that isn't that noticeable by the human eye, but with music everything has to be impeccable and if a note is slightly off, you instantly hear it. Machine learning models that will be able to create outstanding music, imo, will mark t…

> if a note is slightly off, you instantly hear it I don't know about this. I know professional musicians and they say that they make little mistakes all the time, but part of being a professional is being able to pretend those mistakes didn't exist because the audience doesn't know what it's supposed to sound like.

> but part of being a professional is being able to pretend those mistakes didn't exist because the audience doesn't know what it's supposed to sound like.

Very true, however it's something musicians have to specifically train for, because to a trained ear they can become painfully obvious and it's tempting for the musician to (at least) cringe in a way that the audience will detect. Musicians in the audience, though, can often tell when you flub a note, regardless of how you play it off. That's really the bar we are looking to meet.

Re: MusicLM: Generating music from text

#72

Post by spuz »

ben_w wrote: post #56
FeepingCreature wrote: post #23

Weights when though. :-(

For this specific paper, probably never. From the final section: """We acknowledge the risk of potential misappropriation of creative content associated to the use-case. In accordance with responsible model development practices, we conducted a thorough study of memorization, adapting and extending a methodology used in the context of text-based LLMs, focusing on the semantic modeling stage. We found that only a tiny…

Could someone translate this into layman's terms?

Re: MusicLM: Generating music from text

#74

Post by 52-6F-62 »

dariosalvi78 wrote: post #45

AI music is the next thing coming, wait for copyright lawsuits to fall like bombs. I find it fascinating that MidJourney can make a 3D model of my face from a low quality image, rotate it in space, apply it on someone's else body, add coherent shadows and backgrounds, with a very credible result, and yet an AI cannot generate a decent song, which is 1-dimensional and has probably much less internal modelling to care…

> which is 1-dimensional and has probably much less internal modelling to care about

There is an ocean between “I like that sound” and a final, produced piece of recorded music. Much of that ocean being ineffable.

Trying to reduce it to a set of parameters around the final waveform and you’ve missed the entire point.

> but it will suffice for, say, putting a music background for your startup cheap marketing ad.

We need less of that—not more. It’s like a climate disaster of the soul.

Re: MusicLM: Generating music from text

#75

Post by jameshart »

bulbosaur123 wrote: post #62

I'm absolutely flabbergasted by rapid progression of music related machine learning models. I've always seen music as the "final frontier". With images you can have tiny errors and noise that isn't that noticeable by the human eye, but with music everything has to be impeccable and if a note is slightly off, you instantly hear it. Machine learning models that will be able to create outstanding music, imo, will mark t…

The fact that mistakes in music are so obvious comes down the the fact that underneath music is a lot of structure and pattern. That there can be a ‘wrong’ note implies that there is redundancy in the signal and predicting the right next sound is precisely the kind of thing these sorts of generative AIs excel at.

Re: MusicLM: Generating music from text

#76

Post by williamcotton »

As a musician I don’t see these tools as competing with what I do but enabling me to do so much more. It takes a lot of time and money to create a professional recording for my songs and drum machines and synths just don’t work for Americana. These tools offer the possibility of backing tracks that sound like Willie Nelson’s band from the 70s but at a fraction of the time and effort.

I can’t wait until they get to the point where they’re more composable or auto-accompany given an acoustic guitar and vocal input.

Re: MusicLM: Generating music from text

#77

Post by jameshart »

woolion wrote: post #18

I don't really understand why this approached is pushed for music. You can overpaint an image, but you can't do that with a song. Cutting an image to reintroduce coherence is easy too. For a song you need midi, or another symbolic representation. That was the approach of pop2piano (unfortunately it is limited to covers, not generating from scratch). And even if a song generated this is OK, listening to half an hour f…

“For a song you need midi”

Pretty sure the Beatles never handed George Martin any midi files. What’s the symbolic representation that captures the tone of every bend in a Hendrix solo? Did Daft Punk go back and grab the raw master stems of the old vinyl recordings they used to assemble their tracks?

Music producers have been astonishingly creative given inputs in a vast range of formats. Sheet music and midi are one tool, but ultimately it’s about combining sounds in the mix isn’t it?

Re: MusicLM: Generating music from text

#78

Post by williamcotton »

throw_pm23 wrote: post #51
echelon wrote: post #44

Earlier quoted context omitted.

Why is there this need for humans to be involved at all stages of production? Do we still need humans to knit and handwash our clothes? To live in a caboose or lighthouse? To operate our elevators? To deliver the milk? Why should anyone have to learn to draw in the future when machines promise to do a better job? There are so many better uses for our time, and too many things to do for our short lives to handle. I wa…

It may be empowering, but it is not yet clear that the result will be appealing. Not trying to be an AI-skeptic, but I find art to still be mostly about communication; when I listen to some music I enjoy I feel like I have some shared experience with the author which they are able to communicate via music. I have yet to see this effect in AI-produced stuff, but even if the effect would be fully imitated, it is still…

Yes, nothing communicates the lived human experience like chip tune music playing in the background of a video game.

I hear a lot of romantic notions yet I don't see a lot of people flocking to solo acoustic singer-songwriters!

Re: MusicLM: Generating music from text

#79

Post by rrherr »

bulbosaur123 wrote: post #62

I'm absolutely flabbergasted by rapid progression of music related machine learning models. I've always seen music as the "final frontier". With images you can have tiny errors and noise that isn't that noticeable by the human eye, but with music everything has to be impeccable and if a note is slightly off, you instantly hear it. Machine learning models that will be able to create outstanding music, imo, will mark t…

> Anyone want to share their experiments with MusicLM, feel free to join

How can we experiment with MusicLM?

The paper’s last sentence says, “We have no plans to release models at this point.”

Re: MusicLM: Generating music from text

#80

Post by 5040 »

jonas-w wrote: post #25
modeless wrote: post #13

Earlier quoted context omitted.

I love that there's a whole section for accordion examples. I think the accordion rap has a bad word in it. Accordion techno works surprisingly well. They really need to bump up to 48 KHz so all the music doesn't sound like it's being played over a telephone. A factor of two in the cost shouldn't be prohibitive. So much of the audio generation stuff I've seen has fatal flaws like this baked into the dataset and/or tr…

The "problem" is copyright, the music industry has way more power than the people that were copyright infringed by Stable Diffusion et al. And probably no artist that makes a living from their music will voluntarely give his music to AI.

Pirates will gladly provide that content to the AI. Look what's happening with deepfakes.
Post reply on HN