AI Speech Synthesis And The Future Of The Market

BANNER-SIZE-AI-and-Machine-Learning.jpeg
AI Speech Synthesis And The Future Of The Market

A plausible, and bleak future


First of all, I would just like to preface this by stating that this article is based upon a previous post I had made a few weeks back in the Hip-Hop Discussion forum. A lot of good points were made in that thread by various users, whom I will credit when relevant. The thread itself was a small scale discussion, but I feel that the entire LEAKED.CX community needs to have a broader discussion regarding the issue of increasingly complex AI synthesization of speech, especially in the context of this forum/marketplace.

You’re probably familiar with sites such as uberduck.ai

For anybody who isn’t though, uberduck.ai is a website where many celebrity speech models have been generated, and can be interacted with in a text-to-speech manner. The models are hit or miss currently, this is of course depending on how long/well the model has been trained against examples of the vocals it is trying to replicate. These models are not exclusive to these sites however, and anybody with technical experience working with artificial intelligence could feasibly put together a model of their own. This means that in the hands of every human with access to a functioning computer and an internet connection, is a chance for new AI to be created.

Screen Shot 2021-08-25 at 12.13.15 PM.png

The issue facing us is that the legitimacy of audio in and of itself is now being threatened. As you all know, LEAKED.CX is a website founded upon the principle of the legitimacy of sale. This site, being a well known marketplace for what is assumed to be genuine privately owned songs created by celebrity artists, is now in the crosshairs of those who may have the will and technical experience to falsify audio. Not just for fun, but for the purpose of sounding so legitimate that one could be scammed out of hundreds, if not thousands of dollars with no recourse. This is not going to be an isolated issue, and time will only make this issue worse.


The future is now

I am certainly far from the first to consider this as a possible point of failure for the entirety of the unreleased music market. You may think “the examples of AI voice / song replication are rudimentary at best” and broadly speaking, for this moment in time, it is true that the vast majority of these models are unable to deliver a quality of sound that could fool a buyer out of their money. Unfortunately however, AI trends towards progress with time. A properly maintained AI model will never evolve backwards, and in this case, the desired trait to obtain is the accuracy of the audio in comparison to real songs. This process will certainly take time to develop into a serious threat, but I can guarantee you that as you are reading this article, there are models being trained to replicate your favorite artist’s songs; if they aren’t right now, they will be soon. From the outside looking in, it may seem like a very complicated thing to get into, and one may hope that the complexity alone would be enough of a deterrent to ward off scammers, but it is not. Those who understand what they are doing on a technical level, and with a little hope for the future of the technology that is powering the model, see before their eyes a true money printing machine, and the person reading this article is the desired victim.


The depth of the issue

This of course is a thing that not only will influence the future of this website, but the music industry as a whole. There will likely be a day where a label decides to train their own model using raw stems from a deceased artist’s project files. Whether or not a public moral conversation will stop this, it will still likely happen at some point. These industry AIs would be exponentially easier to train, having complete stems, and being able to train with such accuracy could create near identical replications of an artist’s unique vocal inflections and style of speech.

You may be wondering, what makes industry AI an issue for LEAKED.CX?

(thanks to BaphometicOrder for this contribution to the discussion)
As you know, the process of obtaining unreleased music begins with a song being created in a studio with an artist, and then a myriad of methods can be used to get those files into the hands of a seller. In this context, if an industry AI creates songs that end up in the hands of a seller, and the seller is unbeknownst to the fact that they are indeed not legitimate creations, they will still sell them, and charge appropriate prices. Thus circulating what could be considered a scam song, even though nobody could ever have told the difference in the first place.

We may have to rethink what it means to be scammed.

If it is true that this is the future of the market, there will be moments like these where nobody is truly to blame, and yet one could also argue that money was lost on a file that never was truly recorded. However, technically speaking, especially in the context of stolen industry AI songs, they would still be exclusive and presumably of very high quality. As backwards as that may sound at this very moment, to consider something exclusive that never was made by the artist you really wanted, it may be a challenge we will face in the coming years, if this becomes commonplace and the markets are muddied with fake and real songs that are impossible to differentiate between. I’ll let these two quotes speak for themselves:

“is it really a scam if you receive a fully finished unheard song no one would be able to replicate except the person who made it though?”
- LEAKED.CX user Young Thug

“it is the same premise as the artist, the artist made an unheard verse that no one can replicate, and the AI did the same thing”
-LEAKED.CX user BaphometicOrder


That is a very valid argument to make. I think the discussion can reach an even deeper level though: perhaps it would make more sense, if this ever does become a problem, to create a new section of the marketplace, specifically for AI. Assuming it was user made/verified to be inauthentic through means we can discover in the future. Imagine a low price marketplace where the most famous artists in the world can be technically bought for dirt cheap, and you still are the only person with the file. In this marketplace, prices could be determined on how good the songs ended up being after generating them. Imagine being able to buy a $30 “Playboi Carti” song that nobody else will ever hear besides you and the creator, and it sounds completely real. Perhaps this marketplace could have a request section, where you could outline the parameters of the track, AKA, the era, vocal style, and much more to suit exactly what you wanted.

This might even be the deterrent we would need to keep the majority of scams out of the true marketplace. But that doesn't remedy the situation entirely. Of course scams will still be attempted, so perhaps we should preemptively come up with a new set of parameters for the song verification process, which can help identify such fakes. Whether that would be allowing them to be run through an identification application, or providing more information, I'm personally not sure.


In conclusion, we should have this in our minds going forward, or else this site, and the entire unreleased music market as a whole has the potential to implode.

Thank you for your time,

- 03 Greedo
 
Please, Log in or Register to view quote content!
Listen to me very carefully. Ally Lotti held Juice captive in a sweat shop. She restrained him and forced him to pump out hundreds of songs with a hamster’s water dispenser filled with lean beside him that she only dispensed when he finished another song. She let him out solely for live performances to make everything seem normal and keep the dark truth under the radar.

Jarrad has such an amazing work ethic and determination that he still managed to get fat off a lean-only diet despite her strict rationing regime of one dribble per song made.

Ally did this so she could sell the songs behind his back on the black market and give up her previous working life that consisted of smuggling meth across international boarders. She now spends her days role playing as a character from Desperate Housewives. #woke

Mods you can give me author role tomorrow morning if you’re too busy tonight np
 
Amazing article, you’re setting the bar high for future authors. I’m not very familiar with the current speech synthesis tech. Do you think AI is already advanced enough to achieve convincing clones today if people spent more time on trial and error and mixing?
 
  • Like
Reactions: 03 Greedo
gonna be at least 2-3 years before any public algorithms and models are good enough to trick keen listeners. who knows, there might be some massive breakthrough and it could be sooner, but I kinda doubt it since the pace it's been going through the past ~2 years has been pretty consistent, and unless scammers start going out of their way to learn about AI and coding their own tools & training their own models it's gonna be a while. things like uberduck are probably never going to be of concern, none of their models sound very realistic at all, and as long as it's a web app it will probably stay that way
 
Please, Log in or Register to view quote content!
I think it depends on the artist truthfully. Take this for example:
This was made by the site I spoke about in the article, uberduck.ai

I believe because Carti has a generally more simple and less rounded way of rapping, it is easier for AI to replicate his songs in a convincing way at this point.

This song truthfully confused me at first! I didn't know whether it was real or not (posted in leak section) then I was shown this video. It actually was a major inspiration to write this article.
 
  • RT
Reactions: DIG
this is actually a really interesting read and pointed out some stuff i hadn't even thought abt on the topic. well written too good looks
 
  • Like
Reactions: 03 Greedo
One thing the article doesn’t touch on is how the AI is trained. Currently stuff like uberduck is built on released songs, possibly roughly isolated to generate the replica vocals, and this is already enough to be a upcoming major problem in snippet verification but the real issue is if high end hackers or even labels feed similar deepfake/ai trainers the terabytes of pure crisp vocal takes that are privately available.
 
IK for a fact that some members on this site have already been sold AI songs privately. Sellers have already trained much better models than the publicly available one, and anyone with a bit of determination can make better models than the one Isak made for uberduck. This will lead to the end of marketplace.
 
Not really related to leaked.cx community but I really think Kanye could in the future incorporate AI verses in his albums, he really looks interested in technology that allows people to collaborate on the creative process (stem player) and he also used that Robert Kardashian hologram so he definitely must like AI and speech synthesis. Furthermore his KMSA verse on Donda already sounds like AI to me but maybe it's just a weird vocal take.

Imagine if he beefs with Jay just before dropping WTT2 and replaces him with AI-generated verses
 
  • RT
Reactions: 03 Greedo
Top Bottom