Google Created an AI That Can Generate Music From Text Descriptions, But Won't Release It (techcrunch.com) 52
An impressive new AI system from Google can generate music in any genre given a text description. But the company, fearing the risks, has no immediate plans to release it. From a report: Called MusicLM, Google's certainly isn't the first generative AI system for song. There's been other attempts, including Riffusion, an AI that composes music by visualizing it, as well as Dance Diffusion, Google's own AudioML and OpenAI's Jukebox. But owing to technical limitations and limited training data, none have been able to produce songs particularly complex in composition or high-fidelity. MusicLM is perhaps the first that can.
Detailed in an academic paper, MusicLM was trained on a data set of 280,000 hours of music to learn to generate coherent songs for descriptions of -- as the creators put it -- "significant complexity" (e.g. "enchanting jazz song with a memorable saxophone solo and a solo singer" or "Berlin '90s techno with a low bass and strong kick." Its songs, remarkably, sound something like a human artist might compose, albeit not necessarily as inventive or musically cohesive. [...] That's not to suggest MusicLM's flawless -- far from it, truthfully. Some of the samples have a distorted quality to them, an unavoidable side effect of the training process. And while MusicLM can technically generate vocals, including choral harmonies, they leave a lot to be desired. Still, the Google researchers note the many ethical challenges posed by a system like MusicLM, including a tendency to incorporate copyrighted material from training data into the generated songs.
Detailed in an academic paper, MusicLM was trained on a data set of 280,000 hours of music to learn to generate coherent songs for descriptions of -- as the creators put it -- "significant complexity" (e.g. "enchanting jazz song with a memorable saxophone solo and a solo singer" or "Berlin '90s techno with a low bass and strong kick." Its songs, remarkably, sound something like a human artist might compose, albeit not necessarily as inventive or musically cohesive. [...] That's not to suggest MusicLM's flawless -- far from it, truthfully. Some of the samples have a distorted quality to them, an unavoidable side effect of the training process. And while MusicLM can technically generate vocals, including choral harmonies, they leave a lot to be desired. Still, the Google researchers note the many ethical challenges posed by a system like MusicLM, including a tendency to incorporate copyrighted material from training data into the generated songs.
It can't generate good guitar solos (Score:4, Funny)
Re: (Score:2)
You mean like Animusic's [youtube.com] Resonant Chamber [youtube.com] ? =P
Re: (Score:1)
Copied music most likely (Score:1)
Re: Copied music most likely (Score:2)
Re: (Score:2)
Sort of? You might be referring to the fact that they got a court ruling in the U.S. stating that it was not copyright infringement to train your AI on copyrighted material.
I don't think anyone expects though that if your AI content-generator happens to overfit on something and spits out actual copyright infringing content that it's magically copyright-free/available for the person running the generator to copyright (and whether this could ever be a problem for the AI developer, instead of just for the AI *
Re:Copied music most likely (Score:4, Insightful)
Well, that's how the big groups like Zeppelin, the Stones, etc all did things..they "nicked" things here and there from the past and made new music built upon the past...like those earlier groups before them did.
Ok, Zeppelin was a bit more blunt about it that most other groups, but they ALL have done it...that's how music evolved.
Won't release mean aiming for SaaS? (Score:2)
Re: (Score:3)
I think it's far more likely that they realize that MusicLM will open them up to massive copyright infringement suits of the kind Microsoft is facing with Copilot. AI will always have the fatal flaw that it is just sophisticated pattern matching. It is, and always will be, incapable of originality, and can only iterate over known possibilities.
Re: (Score:3)
Yes. I don't know why that is so hard for so many to understand.
There is no creativity here, no aesthetic judgement, it's all just probability. That doesn't mean it can't produce something new, these things obviously can, only that the "new" thing is necessarily similar to the things in the training set.
It's worth pointing out that we've had AI music generators for ages. They're actually really easy to write, and can be a fun project for beginners. They can also be made to make music in any style, or a
Re: (Score:2)
It is, and always will be, incapable of originality, and can only iterate over known possibilities.
What is the difference between humans and any possible future AI that makes humans capable of originality and AI forever incapable of it?
Limitations as described (Score:3)
The limitations listed in terms of complexity, nuance, etc. make it sound at least as good as the writers of "most" available music today. See Rick Beato's video on the vanishing of key changes in modern music as one example.
Great work, however you choose to define it, is a ways off yet, but "good enough" is on the doorstep. If I need generic background music for an ad or a low budget documentary I would absolutely reach for it.
Re:Limitations as described (Score:5, Interesting)
The limitations listed in terms of complexity, nuance, etc. make it sound at least as good as the writers of "most" available music today. See Rick Beato's video on the vanishing of key changes in modern music as one example.
From the description, yes. From actually listening, not so much. Mind you, it's very cool that it can at least demonstrate understanding of what we mean by the text and generate something that technically qualifies as representing that text. But it's a stretch to call the result "music".
Basically, it all falls under the ultra-modern "ambient music" genre, which is to say, music that make musicians appreciate Van Gogh more (and I mean the ear, not the art). I'm not hearing anything I would want to listen to in any of those samples. There's not a recognizable melody, there's not harmony, there's not any passable chord structure (or any other kind of structure, including any sense of rhythm or time signature, generally speaking), nor voice leading, nor anything else that would make it music. It is more like sound sculpting than music, like taking random bits of music and jamming them together nonsensically.
Maybe it will get there eventually, but I kind of have to ask why one would bother. Instead of seeing this sort of tech try to create, which it really isn't very good at (at all), I'd much rather see these techniques used to make it possible for people to realize their visions. Start from a melody and chord structure, and let it come up with various ideas for choral parts, a little bit at a time, so you can choose from various approaches as you go along, and it can adapt to your preferences. Paste in choral parts with words and notes in any language and have it sing them, so that you can hear what something will sound like when sung by a singer or a choir. And so on. Let people create the framework, the structure, the basic idea, and let the technology help turn that into something amazing.
I would kill to be able to paste in a choral score and hear even a semi-passable performance, and that should totally be possible just by training on existing sheet music and recordings.
Re:Limitations as described (Score:4, Interesting)
"Instead of seeing this sort of tech try to create, which it really isn't very good at (at all)"
I think you may be falling into the same box that I've seen in a lot of threads about AI lately - judging the current state instead of the potential. What you're calling "not very good at all" is a vast improvement from "nonexistent" in a very short period of time.
Look at only a few years ago, when attempts to make a robot simply walk were hugely problematic. Now the damn things are getting scarily mobile. ChatPGT might not produce great literature, but it can already pass plagiarism checks, pass university exams, and produce better written work than a lot of students. Music is just another domain, and "not very good at all" today, will mean passable really soon, and mid tier pretty soon after that.
I don't think we're going to see the top tier of musicians and writers replaced any time soon. But those in the middle better pull up their bootstraps, and those at the bottom should maybe re-evaluate their career goals. Making a living writing commodity background music will be gone.
Re: (Score:2)
I don't think we're going to see the top tier of musicians and writers replaced any time soon. But those in the middle better pull up their bootstraps, and those at the bottom should maybe re-evaluate their career goals. Making a living writing commodity background music will be gone.
I thought it already was. The sheer volume of free background music out there is so large that I can't imagine anybody making a living off of it. :-)
Re: (Score:3)
"Instead of seeing this sort of tech try to create, which it really isn't very good at (at all)"
I think you may be falling into the same box that I've seen in a lot of threads about AI lately - judging the current state instead of the potential. What you're calling "not very good at all" is a vast improvement from "nonexistent" in a very short period of time.
Actually, that's not true. Dr. David Cope [ucsc.edu] used neural networks and other similar techniques to create computer-generated music way back in the 1990s, IIRC, or at least the early part of the 2000s. And having played some of that generated music (IIRC, it was an attempt to imitate Mozart, trained on a giant pile of his works) about two decades ago, I can say that this doesn't seem significantly better. If anything, those earlier experiments were more musical, because at least they had some structure to the
Re: (Score:2)
...but rather because I've watched people try to do this over and over again, and have yet to see the quality bar move significantly towards something worth listening to.
This is exactly the perspective from which I've approached my evaluation of AI. I've seen stuff like this since (at the latest) the early 1990's. The largest improvement I've seen is in the hardware necessary to run the software. The access to huge datasets via the Internet has been the second-largest improvement.
I was reading about how neural networks were going to send AI into warp speed since I was in highschool in the mid 1980's. While there have been incremental improvements, they've been barely signi
Re: (Score:2)
...but rather because I've watched people try to do this over and over again, and have yet to see the quality bar move significantly towards something worth listening to.
This is exactly the perspective from which I've approached my evaluation of AI. I've seen stuff like this since (at the latest) the early 1990's. The largest improvement I've seen is in the hardware necessary to run the software. The access to huge datasets via the Internet has been the second-largest improvement.
I was reading about how neural networks were going to send AI into warp speed since I was in highschool in the mid 1980's. While there have been incremental improvements, they've been barely significant.
I wouldn't go that far. In some spaces, they've improved dramatically. Image recognition comes to mind. And even generative art is rather amazing as long as you don't care about realism. What makes music challenging is that realism is such an important part of it. Singers don't glissando wildly from note to note, typically. Real pianos are physically incapable of pitch bend. Yet we hear those sorts of artifacts, which makes it obvious that the performance is something other than plausible.
The earlier
Re: (Score:2)
"I'm not hearing anything I would want to listen to in any of those samples."
That's like 90+% of highly popular "music" promoted today. "Machine generated" is hardly different from the formula-generated crap aside from how much involvement there is from some nonmusician operating the controls.
Simplistic tunes (Score:2)
Listening to the MusicLM output and speaking as a composer, this only demonstrates that 99% of the music created by humans is no better than what a computer can generate.
Re: (Score:2)
Give this [youtube.com] to the A.I. to see if it will improve its sarcasm or completely destroy it.
Re: (Score:2)
Re: (Score:2)
Wait... do you you mean coke?
why not releasing is obvious (Score:5, Insightful)
"including a tendency to incorporate copyrighted material from training data into the generated songs"
Unlike text and image copyright violations, music publishers almost always win copyright lawsuits even for the tiniest extract of sound or melody that they think some jury or judge would recognize.
Re: (Score:1)
Heh, I guess they're "under pressure" not to release models for music as a result :)
Re: (Score:2)
The bad timeline (Score:5, Interesting)
Re: (Score:2)
they can already draw, make music, write stories or code.
Not really.
Re: (Score:3)
Re: (Score:2)
Re: (Score:2)
Re: (Score:2)
this tech is improving at a crazy fast rate.
I wouldn't say that. The technology is essentially the same as it was 20 years ago. The only real difference here is resources (memory/processing) and massive amounts of data.
If you want to see actual improvement, that'll take a fundamentally different approach. We're in a boom now. Expect a bust when reality catches up the the hype.
Re: (Score:2)
Information workers being replaced by information synthesizing machines that are a product of the information age? Who saw it coming?
Me. The major issue is when an art robot screws up and makes seven fingered people, no one dies, nothing is destroyed. When a self driving vehicle screws up, people die and things are destroyed. That's true of all the blue collar fields: it is tough work because it's dangerous, and screw ups get people killed in equipment destroyed. Not true of any art or creative product. Eve
Re: (Score:2)
"It seems AI will soon allow humans to stop spending time on creative activities and focus on boring hard physical tasks..."
I tried to think of something pithy to say about this, but all I can come up with is, "Well, shit."
Re: The bad timeline (Score:2)
Re: The bad timeline (Score:2)
Re: (Score:2)
I wonder (Score:2)
Re: (Score:3)
Re: (Score:2)
I tried it. A weird voice chanted: "Must cancel popular projects. Must cancel popular projects..."
That is backwards, because Google cancels unpopular projects. Anything with a few hundred million users is safe.
If its not public, Why do we care? (Score:2)
We already knew this was possible, and other people are currently working on this themselves publicly as open source projects. So why do we care that Google is doing this if they're not going to release it? Why is this even posted on slashdot?
I'll tell you why.
Because Google is scared, and this is a slashdotvertisment to help us all think that they haven't lost it, to tell us that they are working on it now and have done stuff so shouldn't discount them, its a fluff puce to make the stock not drop like a st
Re: (Score:2)
Google's trying to parade it as something that gives them value to us while refusing to give us the actual value.
They released the paper, "anyone" can implement it and do research. The actual value is given. OTOH releasing a website without the paper would have been the commercial move you described.
I can make random music too! (Score:2)
I just smash the keyboard a bunch of times. I dare you to tell me that's not music.
Won't Release It? (Score:2)
Lyrics generated:
Please release me, let me go
...
For I don't love you anymore
Stable Diffusion can already do it (Score:2)
FYI, you can already create spectrographs in stable diffusion that can be converted to music.
Prisencolinensinainciusol (Score:2)
And have GPT fix the lyrics (Score:2)
If AI produces funk rap with gibberish lyrics, AI has a fix for that too. Pipe the metrical structure of the piece through a large language model such as GPT-3 to generate plausible lyrics, and render those.
It won't release it because... (Score:2)
Stuff like this... (Score:2)
is why Google is getting kicked in the teeth, and are in red alert mode, in regards to ChatGPT. Instead of being at the head of the curve, they find themselves playing catchup.