Open-sourcing AudioCraft: Generative AI for audio • Hacker News Board

Re: Open-sourcing AudioCraft: Generative AI for audio

#301

Post by smoldesu » Thu, Aug 03, 2023, 12:05 AM UTC

lannisterstark wrote: post #292

smoldesu wrote: post #289

Earlier quoted context omitted.

I'm able to do everything in your hypothetical scenario besides illegally copy and redistribute the original content. If it makes me happy to write my name on the cover and pretend I wrote "Robert Frost's Poetry Collection" then so be it. Truly, if I want to say "fuck the original artists" in the comfort of my home or margins of the pages, I can do so. I can even sell that adulterated copy under the First Sale doctri…

>that content to be redistributed against your will That is fine. >If it makes me happy to write my name on the cover and pretend I wrote "Robert Frost's Poetry Collection" then so be it. You can do whatever you want in your own house. No one is trying to dictate terms of what you do in your own private time in your own domicile. The problem occurs when you try to profit from my work, by pretending you did the work,…

> However, the idea of "You can add a chapter to this, call it your own, and sell it" and I won't legally come after you is absurd.

The only part of this that seems legally problematic is copying and redistributing the part that you wrote. If I wanted to buy 10,000 copies of A Game of Thrones so I could staple my fanfiction to the back and sell it on Etsy, there is no legal precedent suggesting I could be stopped. It is a lawful transformation of a legally licensed product, payed for in-full and redistributed in accordance with it's individual license. Absent any extenuating contracts between the owner and seller, I don't see what legal ground you would have to stand on.

> and if the results of said legal pursuit result in you being bankrupt...what was the phrase you used, "then so be it."

"if" indeed. Let's check in with the Author's Guild and see how this fight is going: https://www.copyright.gov/fair-use/summaries/authorsguild-go...

Re: Open-sourcing AudioCraft: Generative AI for audio

#302

Post by PaulDavisThe1st » Thu, Aug 03, 2023, 12:10 AM UTC

cmgbhm wrote: post #241

PaulDavisThe1st wrote: post #92

Earlier quoted context omitted.

Establishing derivation is at the crux of all legal matters surrounding diffusion models. It has not yet been clearly established. If it is, then I'd agree with you. Until then, I think it's a bit more up in the air. Also, IIRC, RIAA did not bring many resources to bear against e.g. "home taping" itself, because they could essentially never know that it had occured. The overwhelming majority of their efforts went int…

https://en.wikipedia.org/wiki/Audio_Home_Recording_Act https://en.wikipedia.org/wiki/Home_Taping_Is_Killing_Music Recording industries have fought end user reproduction often. They’ve fought sampling battles. Go after the pocketbooks and go after the technology waves. If there’s a derivative argument they can make, they will.

They may have gotten the AHRA passed, but they essentially lost in every important way:

> "This exception was crucial in RIAA v. Diamond Multimedia Systems, Inc.,[14] the only case in which the AHRA's provisions have been examined by the federal courts. The RIAA filed suit to enjoin the manufacture and distribution of the Rio PMP300, one of the first portable MP3 players, because it did not include the SCMS copy protection required by the act, and Diamond did not intend to pay royalties. The 9th Circuit, affirming the earlier District Court ruling in favor of Diamond Multimedia,[15] ruled that the "digital music recording" for the purposes of the act was not intended to include songs fixed on computer hard drives. The court also held that the Rio was not a digital audio recording device for the purposes of the AHRA, because 1) the Rio reproduced files from computer hard drives, which were specifically exempted from the SCMS and Royalty payments under the act, 2) could not directly record from the radio or other transmissions. "

From the AHRA itself:

> No action may be brought under this title alleging infringement of copyright based on the manufacture, importation, or distribution of a digital audio recording device, a digital audio recording medium, an analog recording device, or an analog recording medium, or based on the noncommercial use by a consumer of such a device or medium for making digital musical recordings or analog musical recordings.

and from Wikipedia again:

> In regard to home taping, the provision broadly permits noncommercial, private recording to analog devices and media. However, it fails to resolve the home taping debate "conclusively," as it only permits noncommercial, private recording to digital devices and media when certain technology is used.

> Two reports by the House of Representatives characterize the provision as legalizing digital home copying to the same degree as analog. One states "in the case of home taping, the exemption protects all noncommercial copying by consumers of digital and analog recordings,"[22] and the other states "In short, the reported legislation [Section 1008] would clearly establish that consumers cannot be sued for making analog or digital audio copies for private noncommercial use."[23]

Similarly, language in the RIAA v. Diamond Multimedia decision suggests a broader reading of the Section 1008 exemptions, providing blanket protection for "all noncommercial copying by consumers of digital and analog musical recordings" and equating the spaceshifting of audio with the fair use protections afforded home video recordings in Sony v. Universal Studios:

>> In fact, the Rio's operation is entirely consistent with the Act's main purpose – the facilitation of personal use.

Re: Open-sourcing AudioCraft: Generative AI for audio

#303

Post by bottlepalm » Thu, Aug 03, 2023, 12:16 AM UTC

squidsoup wrote: post #274

famouswaffles wrote: post #122

Earlier quoted context omitted.

>and AI not yet capable of matching even average human content (much less pro level) Yeah this is not true. Sota Text, Image generation is well above average baselines. You can certainly generate professional level art on Midjourney

Commercial art and Art are not the same thing.

Unless you have a tool that can tell the difference, they are.

Re: Open-sourcing AudioCraft: Generative AI for audio

#304

Post by polygamous_bat » Thu, Aug 03, 2023, 12:17 AM UTC

__loam wrote: post #210

smoldesu wrote: post #89

Earlier quoted context omitted.

That, or this is an ongoing research lab (FAIR) that has existed for ~half a decade and has advanced the state-of-the-art in AI further than Apple, Microsoft and Google combined.

I would be pretty shocked if meta were that far ahead of all 3 of those companies, all of which are also spending a fuck load on internal AI research.

> all of which are also spending a fuck load on internal AI research.

But their internal research stays internal. Sometimes, they put out "papers" which are glorified advertisements, often going as far as hiding the model architecture just to keep their competitive advantage.

Re: Open-sourcing AudioCraft: Generative AI for audio

#305

Post by goosinmouse » Thu, Aug 03, 2023, 1:35 AM UTC

parhamn wrote: post #225

javajosh wrote: post #211

Earlier quoted context omitted.

You're not supposed to actually install it and use it, just comment on how cool and open Facebook is, especially in comparison to OpenAI. So, user error.

Right its not like anyone has operationalized Llama 2 or there aren't hundreds of repos for inference servers and the likes. /s

I just made a script that generates a two minute long classic radio show episode in the style of Johnny Dollar. Im using elevenlabs for the dialogue so the audiocraft element is definitely the weak point, but its super neat that its even possible currently.

Re: Open-sourcing AudioCraft: Generative AI for audio

#306

Post by manquer » Thu, Aug 03, 2023, 1:56 AM UTC

PaulDavisThe1st wrote: post #92

manquer wrote: post #68

Earlier quoted context omitted.

Even before streaming, You never "owned" any music legally [1], you merely owned a physical copy of a performance[3] of a song, that in no way gives you the right to make derivative works [2] automatically. Also it doesn't really matter on what the law says, RIAA in the last iteration, relied on the fact it you would rather pay a fine than be able to afford expensive lawyers to fight the specifics out in court on ave…

Establishing derivation is at the crux of all legal matters surrounding diffusion models. It has not yet been clearly established. If it is, then I'd agree with you. Until then, I think it's a bit more up in the air. Also, IIRC, RIAA did not bring many resources to bear against e.g. "home taping" itself, because they could essentially never know that it had occured. The overwhelming majority of their efforts went int…

RIAA and others in general sued a lot of people including bar owners for playing their songs etc , there were also some enforcement via private companies with three strike policies and so on especially in Europe .

They went after ISPs and torrent sites which only hosted magnet links and many others who shouldn’t have been really sued

The goal was to create a very hostile environment for downloading songs to protect their interests - “you wouldn’t download a car!”

It was never the goal nor ever realistic to actually pursue enforcement action against every offender , the idea was to change behavior with all the related actions .

They did end up changing behavior, people just didn’t want the hassle, or be in fear so paying for streaming for access had a stronger value proposition, it was not what the RIAA planned , but benefits enormously today.

Re: Open-sourcing AudioCraft: Generative AI for audio

#307

Post by dragonwriter » Thu, Aug 03, 2023, 2:59 AM UTC

vel0city wrote: post #269

dragonwriter wrote: post #128

Earlier quoted context omitted.

> “Noncommercial art” is not a thing in the eyes of the law The definition of “NonCommercial”, the oddly capitalized term of art in the license, is not a matter of general law, it is a matter of the license, which defines it as “not primarily intended for or directed towards commercial advantage or monetary compensation. For purposes of this Public License, the exchange of the Licensed Material for other material sub…

In their example audio clips they have a "perfect for the beach" audio track. With your understanding of the NC license, would a resort or private beach club be able to play a similar generated music track at their poolside bar or something along those lines? Their primary intention of the bar isn't to play the music, its just an additional ambiance thing; they're trying to sell drinks and have guests pay membership…

> With your understanding of the NC license, would a resort or private beach club be able to play a similar generated music track at their poolside bar or something along those lines?

A resort, probably not, ambiance is, at least arguably, a marketable commercial advantage; a private club in the “mutual benefit organization” sense (rather than a “business selling memberships”, which is just like a resort), probably, because their interest, even indirectly, isn’t making money.

Re: Open-sourcing AudioCraft: Generative AI for audio

#308

Post by squeaky-clean » Thu, Aug 03, 2023, 4:37 AM UTC

jeffbee wrote: post #154

TrackerFF wrote: post #133

Earlier quoted context omitted.

I've played guitar 25 years, and it's funny how the music community has been using all kind of words to describe music or tone. Describing certain tone as "hairy", "mushy", "wooly", "airy", "buttery", etc. is just very common.

Sure, there's jargon. But these words don't describe the music. They describe words associated with the kinds of people who would listen to it (according to the biases of the language model). As a description of the music, it's meaningless. If a person was asked to name some "environmentally conscious" music they could just as easily veer over to hardcore straight edge.

It works if you don't need a specific type of music but just something to fit a vibe. For example in an indie game or film project. Maybe level 3 is a jungle area with a plant based species of NPCs and the music should be "Earthy tones, environmentally conscious ... organic instrumentation"

Re: Open-sourcing AudioCraft: Generative AI for audio

#309

Post by e1ghtSpace » Thu, Aug 03, 2023, 4:38 AM UTC

makestuff wrote: post #286

These models are going to end up being used for advertising. Soon pretty much every ad you see will be generative AI based. It makes A/B testing way easier as you no longer need a creative person to modify the ad or change something subtle about it. For example, the generative voice might change to a different speaker or something, and the AI can generate thousands of different voices to see which one is most effecti…

We might even end up with the most effective ones being the weirdest. Where people clicked though because the ad was so weird.

Re: Open-sourcing AudioCraft: Generative AI for audio

#310

Post by cooper_ganglia » Thu, Aug 03, 2023, 5:39 AM UTC

crazygringo wrote: post #131

ysleepy wrote: post #126

Earlier quoted context omitted.

I don't agree. With ML tools it is possible to make sweeping changes to images and text that are often impossible to detect. combined with the centralisation of most online activities, large players could alter the past. Imagine facebook decides to subtly change every public post and comment to show some particular person or cause in a better light.

If one "large player" like the NYT decides to "alter the past", you can compare with the WaPo or any other newspaper. You can compare with the Internet Archive. You can compare with microfiche. These aren't "impossible to detect", they're trivial to detect if you bother to compare. We have tons of credible archived sources owned by different institutions. And these sources are successful in large part due to their cr…

Your solution, in the case of trusted sources altering content to fit a particular worldview, is to look at other "trusted" sources. I think that therein lies the problem. I believe the real danger isn't people being convinced of something untrue. I think the real danger is the apathy that builds up as people can no longer reliably distinguish the truth, and they give up on sifting through it altogether, instead accepting "their truth". The vast majority of people simply don't care enough to verify sources.

This is already happening without generative AI, and this new stuff is only going to speed things up exponentially.