The Illiac Suite - Music and AI

This is an automated AI transcript. Please forgive the mistakes!

Hello humans, I am angry. And the more I talk about AI,
the more I get angry. Seriously, what is going on? I can see around me a lot of
people in the art world already struggling because people are not willing to pay
them anymore when they can just ask an AI to do the same work. The danger is
real. And the tale of the big tech companies that AI will make people work less
for the same payment is just bullshit. Excuse my words. One side of me just loves
what is possible right now with AI in the music, amazing. It is the reason why I
do this podcast. The other side, a couple of weeks ago, Suno posted. We are very
excited to announce that throughout the remainder of 2024, we will be paying a total
of $1 million to these creators who are bringing us closer to music and each other
with Suno. We are kicking things off with the summer of Suno, a celebration of the
songs that move us as a community. Whether they make us bob our heads, tap our
feet, laugh, cry, or simply say wow. - Seriously, you are paying the artists who use
the music you have stolen from a million of other artists before paying the artist
you have stolen from? I would love to play "Ironic" now from Alainis Morissette
here, but I don't have the rights, so you will hear me sing maybe the only time
in the Iliac Suite. Isn't it ironic? Let's get this episode started.
This is the Iliac Suite, a podcast on AI -driven music. Join me as we dive into
the ever -evolving world of AI and music, where algorithms become the composers and
machines become the virtuosos. Yes, this music and text was written by a computer,
and I am not real, but-- - I am and my name is Dennis Kastrup.
A lot of things are happening these days again in the field of generative AI,
Sunoo and Udyu have announced their audio to audio tools, the web was flooded with
videos of that. Universal Music just announced that they will work together with
Soundlab. Soundlab can create then a clone of the voices in their roster and then
the artists can decide what they can do with this or what they want to do with
this. Sounds a lot like VoiceSwap a platform that also has AI cloned voices and you
can use a voice to create some other music and you have to pay for this. If you
want to know more about this you can listen to episode 7 here in the Iliac Suite
because I already talked about this and what VoiceSwap did is kind of the same
thing I think what Universal Music does and it goes always like this. The little
platforms do it first and then the big one take the ideas. And then there is of
course the video production dream machine from Luma Labs or Kling people are just
amazed and amazed and amazed and amazed how great that is Memes which were taken as
photos from real life become no real again same with album covers I've seen some
really funny album covers from Nirvana from Pink Floyd. It's really great. So Photos
become real again. so photos taken over the reality become real again full circle I
would say. But we will not talk about videos today. By the way, a little bit.
Have you seen the first ever fully created video with an AI for the band Washed
Out? Go have a look because Washed Out is also one of my favorite bands. Yes,
it's fine. Song is fine. But I would say Room for Improvement. improvement. Okay,
what to expect today in the Alex suite? A couple of weeks ago, I talked to my
friend Matthias, who is by the way running Music Tech Germany, an organization that
takes care, as the name says, of companies that work with music and new
technologies. Check it out, Music Tech Germany. He told me there is an AI platform
called Delphos. It is amazing because it does what we have been talking about
before. So Matthias and I. One day each one of us will have his or her personal
AI with which we can create music in our style and we do not have to infringe
copyrights because it is only trained on your own music. I thought instantly I have
to talk to the person behind that platform. I want to know more. So meet Ilya
Tolchenov. I'm the CEO and co -founder of Delfast Music, which is a AI music
platform that is based in London.
By training I,
at one point I was a mathematician, I thought I'd actually go into that academia
and do all that kind of stuff, but then I decided I would go and be a musician
and I worked as a singer for a while. I played different instruments as well but
singing became like the thing I was really passionate about and I've worked across
church choirs across the UK and then I started developing music tech with my co
-founder who's my dad Roman and we first built like a small app that could generate
sheet
Um, because we've always had this idea of being able to sculpt music with like,
with almost with your hands, not having to put notes on a page, note by note and
get lost in the kind of tedium of that. So we were always focused on ways of
making music, um, more intuitive and allow people to concentrate on creativity.
I, We started Delphos in 2021 and the idea was actually quite loose to begin with.
We wanted to create a set of tools so anyone could create music.
But that developed quite quickly into a generative product because we could see just
how excited people were about customizing the music they make and the music they
listen to, we saw that with Web3. So not necessarily the things that really got
people into Web3, like the economics and the speculation and the potential payday at
the end of it. We just really liked how when people got attached to a project,
they really wanted to customize it and see customization at scale. And we thought
this needs to happen with music. And around this time, by coincidence,
other people were really, really starting to come out with very big AI music
projects. There'd been companies before that, of course.
But now is when the really bigger companies were starting to release their own
models.
I've mentioned this before here quite often. What sticks out with a lot of founders
of music AI platforms is the fact that they all share a deep love for music.
Same with Ilya and his partner, his father. It is a family business, a family
affair. When all these new platforms emerged, the two saw the need for something
else, also having in mind that people were a bit afraid of what was coming. We
know that music has a uniquely human element and there will be people who say
everything will be replaced by virtual artists etc. There are a lot of worries about
that with each stage of technology when the radio became a thing like how are we
going to enjoy music if it's just coming out of a wireless and not in real life
and then when sampling became a thing and when a lot of electronic music became a
thing it was weird for people and I think it's the same thing with AI but we set
out to do this thing where a single musician can have their own model.
That's just theirs. They train on just their own music. And once it's trained,
it can output new music in their style. And that's all it does. It doesn't learn
someone else's style or use someone else's style. It only uses your music and it
only sounds like you.
Because number one, people could use that in their own music making and instead of
when they're starting a new track instead of writing everything in note by note
again they can have the core elements of that track already written in a way they
would have already written it but more exciting me they can put this to work in an
automated way so if you make a sound world which is what we call these models a
sound world and
In the kind of pre AI days, I said to you hi Dennis. I want you to write me a
baseline. You'd say, okay, sure Here's an invoice and wait a week and I'll get it
to you The idea is once you have your sound mold up, I can prompt it myself ask
for a baseline with my
Restraints, so my BPM my tempo, my chord progression,
my key, my song structure, and will generate from your sound world, exactly as you
write it, you get paid and you didn't have to do anything.
And what we've been building for the last almost three years has been this platform
where musicians, producers, catalogs can create these models based on just their music
and share it with other people so people can create their own new compositions using
musical material from producers and composers they love already.
I love the idea. Ilya mentioned that quickly the sound world, let's get into that a
little bit more. A sound world is a private music AI generator that writes music in
one particular style. So it's a fully trained model that comes from a single artist
or a single group of music. Now, what it does is it allows you to prompt exact
things from it. So if you make a sound world, anybody on the planet that you give
access to could say to it write me eight bars of this music at
120 beats per minute, and I wanted to have the chord progression G E minor a minor
D in the key of G and These are the instrumentation options from the soundboard
that I want to use so knowing all of that knowledge that is gleaned from the
original composer, it will write you exactly what you asked for. Those eight bars.
So if you sum it up, Delphos is offering special skills that only you are capable
of doing. And in this case, for example, producing songs. But it is like having
many hands or brains at the same time. So Delphos offers your skills to a lot of
people, maybe 100 or 1000 people at the same time, normally you could do this at
once because you're not able to do this. But instead someone says now all right,
Delphos says then we just take your work skills and use them. And Delphos takes
care that you get paid for each user that uses your skills. And you also have the
control of the data which is used to train the models. Sounds great to me. Music
is is unlike other data and it has so many more different cross sections than you
usually get in large data sets.
There's like say in the world of jazz, there's a million people who would do this
particular subgenre of a subgenre of jazz, but even within them there's something
that takes Donald Lambert, him, and Art Tatum, him, and Fatswallow, Fatswallow are
being able to capture that being able to understand that like musicality and
composition kind of need to be at the heart of it when you when you start building
these things, otherwise you're kind of stabbing around sort of in the dark, trying
to see if this model or this particular algorithm will
With them with the music product It's very very important that the people designing
it have a very clear vision of how it's going to be used and oftentimes that comes
from the From them actually having a real use case for it. So can they themselves
imagine? Opening up their own program and solving a problem that they have that's
not necessarily what you need So if you if So if you just understand that market
very well, but you might not need the product yourself, it doesn't mean you can't
build a fantastic business with it. But with such a new technology, where the
ability to generate new music gives you unlimited scope to create whatever you want.
You can create new hit records. You can dilute someone's catalog. You can
Totally disrupt the samples industry. You can create educational products. There's so
many things you could do with that I think it's very important for for people to
come into creating it already knowing Where that's going?
By the way, all the music you're listening in this episode of the Eliak Suite comes
from Delphos from their sound worlds. Thanks Eliak for giving me some examples of
your music.
Let's I want to be part of the ELFOS and I am a musician. For training big
models, millions of songs are used from many different genres and musicians and
bands. What do I have to feed the AI with so that a satisfying result comes out?
- So a sound world product at the moment is an enterprise product.
And to train It's, you need MIDI. So there are ways.
So if you have the stems for a set of works, you can use the stems to train it,
but the core functionality behind it is MIDI. So you would feed in,
so you would analyze like a set of tracks that you need. So maybe between 50 and
100 tracks in a particular style that you own the rights to. And that's a very
interesting topic in AI at the moment, what you use to train a sound world. So 50
to 100 tracks would make a very good sound world and you feed in the stems to a
sound world and you say here are the bass stems, here are the drum stems, here are
melodic stems, here are harmony stems, here are effect stems and from that moment on
the sound world How do you write the melodies? How do you harmonize things?
How do you write basslines? What do you do rhythmically so that when you run the
sound with new constraints, it can stick to your style while creating something
completely new? What is interesting with this approach, not a full song is used for
training, but the different stems of it. For me, that makes actually sense because
like this, your different skills can be trained better. It's more clear. It is not
the whole mix of everything, but the clear sound of each stem.
style, that style. And you start to lose certain things. And that's when you get
music that doesn't why sounds authentic to that original to what you're going for.
But there's an even bigger problem. That requires a huge amount of different writers
because no human being can just write you a million tracks. So a million tracks
will have thousands of different creators, tens of thousands of creators, you lose
any sense of copyright ownership through that because it's all put into this black
box. You don't know exactly how it works. You don't know which data item influence
that particular thing. I know there are companies that are working on this at the
moment, but there is no, just how complex these models are. It's so difficult to
find who's responsible for this particular thing.
So for those reasons, we decided to take like a much more difficult road and come
up with a totally new model that would allow you to do precisely this.
This actually brings me to kind of the real motivation for it.
One of the scariest things about AI systems is that it takes agency away from the
original musical creators. So people are talking about are these things going to
replace musicians and producers? Is it going to steal their jobs? Is it going to
dilute their copyright? Are they going to lose revenues from it? And people are
finding really like very interesting solutions So these things, like there are these
wonderful certification programs that are coming out now. There is a lot of talk of
licensing, training data, all the rest of it. And I think that's like a very, very
good thing to do.
The fact that some people are having the very conversation of do we reimburse music
creators for using that thing to build a product that's obviously competitive with
them. That's crazy that there's still a conversation around them. But I think we
need to take that a step further and actually give the original creators of the
music the agency to control over how the AI works. Because if you're a musician,
you sign an agreement and you get paid some amount of money and a company trains
on your work, yes, that's probably very fair. But if at a certain point, you say,
I don't want my music to be further diluted by AI. I want out of this thing.
There is no kind of solution for that. You can't just go to, I don't know, like
especially the larger models and say, Hey, can you untrade my music out of this?
Because I'm done. There's no GDPR for training. And, you know, you know how long it
took to get the EU to do something about data with AI, it's it's gonna be,
and by that point, all the models will exist, everyone will be diluted, et cetera.
So, from that perspective, what we wanted was for somebody to not have to opt in
or opt out of something, but they create their own model, they become the official
source of their music. And if at any point they want to say,
my My music, my generative music, won't be used in this particular application or
maybe at all. They can just disable their sound world or delete it entirely and
that's like a very central principle of what we do. Total control stays with the
musicians. I love it. Let's get started now. And I gave an example and I was a
little bit optimistic with Ilya. I said Bob Bob Dylan and Adele would have their
own soundworlds, how would that work?
You can absolutely make them work together. We're actually developing tech as well to
make these elements like play much nicer with each other so so that it becomes like
a very Natural sounding composition, but again that that's the kind of exciting thing
for catalog owners who wants sort of you know big -time customization user generated
content is has been proven to be like a really huge revenue driver in the space
and if they're able to facilitate fully cleared,
fully cleared for distribution,
collaborations like that, without actually having to bother Bob Dylan or Adele, I
mean everybody wins and no one's copyright gets infringed, more copyright gets created
and everything's done faster.
Everybody wins until the moment of creation, but the songs have to be licensed then.
Who gets the copyrights? So, we have two sides to this. There's one side where
people who make sound worlds monetize their AI on their own,
and then there's the side where we monetize it. For people who choose to monetize
on their own, we have software licensing, so people can license sound worlds, they
can train and retrain them as much as they like, and create music with them. And
they can literally just use it to create more catalog. So, you know,
they license sound world, they pay per use. So if they want to write a million new
tracks, they pay for a million new tracks using the API. But this same API can be
plugged into products and unfortunately, I can't speak right now about some of these
products, but there are some extremely cool things that are going to be released
over the summer and it's a key three of this year of people who've made their own
generators. So that are very, very specific to their sound. Some of these are so
high quality
Um, it's actually quite crazy how authentic they sound man, but you know, you would
go to somebody who typically just releases music and
Just just use their generator straight from them So our models there like are kind
of a typical SAS model with with some strategic partnerships There are revenue
sharing kind of things as well
which I think like in an emerging industry like AI music is quite important that
good players who are on the right side of protecting artists' rights and everything
work together on making an economy that works for the artists and for the end
users. On the other side, we're working on a series of products that come from
Delphos directly for end users. So we are working on a standalone app that functions
very similar to VST so people can publish their soundworld onto this app and then
anybody, anybody, will be able to access it and generate music in their DAW so the
app itself will be free and we'll have some of our own us native models on there
so people can generate you know whatever they need to then we'll start have branded
soundworlds so if a label releases a soundworld so for example and we've had this
with some labels where their business model doesn't allow them necessarily to create
like a software product like that They focus on more traditional ways of making
music, but they have really, really cool music, really cool data that can be
leveraged. So that's us monetizing it. So we'd say just give us your data, we'll
spin up this model, no charge, and we'll deploy this. So we're working on a few
integrations with the DAWs at the moment, which is very cool because then all of
the users of those DAWs will be able to get, for example, the Bob Dylan sound
world and then Bob Dylan's old stems are generating kind of new revenue that just
didn't exist before and the artist is cut into that revenue. One thing that is an
ongoing problem or let's say issue these days is the fact that platforms have to
face the risks of people using AI generated songs so they can pretend to be a real
musician. They can generate something and say, "Well, that's me. I'm a musician."
Spotify knows what I'm talking about, but there might also be a problem of people
pretending to be a musician. They are not. I mean, I could, for example, upload
music from someone else, so I could pretend to be someone else. Anything that
involves user -generated content runs that risk of people,
you know, getting a sound world and then training on Adele's music when they're not
Adele themselves. There are like, because they're still inputting music,
there is a lot of anti copyright infringement that can be done with that,
I think what would actually be even better. And this is a project we've been
working on Um, for a while and we, we'd be keeping under wraps.
We're probably going to launch it at the same time as, um, our wider release is a
way for people to self satisfy that the way they've trained their own personal sound
world has been ethical. So it's, it's come from their own music.
And the way you need to do that is people need to be able to show their own data
set that they use without giving it away, and for anybody to be able to go in and
check it. That's the kind of antithesis to certification programs, which I think are
very, very necessary. But the more people get certified on a particular program,
the more risk there is. next person who kind of comes in and somehow slips through
the cracks, they end up screwing everyone else who's already on it. And there's so
much more pressure on that certification program
to do this, although I know that's a whole debate and I don't want to open a
whole kind of worms on that, but I think ultimately people need to be able to
publicly stake that thing. Because at the end of the day, if you put up an Adele
song illegally on Spotify, because it's come from you through the distributor,
you can be hunted down and made to pay for that. With Soundworlds, there's the
exact same thing. So there's a lot of reporting that we built into the system. So
if, for example, you make a Dennis Soundworld, and you say, actually,
this isn't going to be royalty free. This is going to be a very royalty not free
sound world. And I generate something from it in my DAW. You will know through the
system on your dashboard that I, Ilya, generated this exact bit of music from your
sound world at this point. So if I then go off and have a hit record,
and you know it sounds like that you have like time -stamped evidence to say yep no
Ilya owes me a bunch of money which is words I never want to hear well as I'm
not really a good musician or to say it better and honestly I'm not a musician at
all I cannot play an instrument well I used to play when I was younger but I just
quit that's just between you and me so I am not a musician Don't worry, Ilya, you
cannot take my music and make a hit with it. On the other side, in times of AI,
am I considered a musician if I generate stuff? And let's say use a DAW to mix
the stamps I have generated. Maybe in the future, questions will come up. At what
point are you a musician? I would agree that prompting something and letting AI sing
for you is not enough or is it? Maybe? Maybe you write good lyrics.
But let's imagine I created some stems and just put them together and use an AI
voice to sing it. Am I a musician? What will define a musician? Maybe the same
question exists when it comes to drum machine and auto tune. Are you a drummer and
singer when you use those two things? Let me So what you think, you can always
write me an email to mail @deniscastrup .com, mail @deniscastrup .com.
My contact details will also be in the notes to this episode, also all the
information and links from Delphos, whose users are, by the way. Our users are very
predominantly professional users, so we have a few record labels using this for
production, for remixing, we have music creation platforms integrating together with
the API. We have some catalogs that are building new catalog using it.
So yes, this is very much aimed at the professional market.
That question answered makes me think of one unanswered one. Remember in the
beginning of this But when Ilja said he and his father are running Delphos, alright
Ilja, how do you two work with each other? How is it working? I've always, my
entire life, I've called him by his first name. Not because of any like,
relationship thing, it's just like, he's just been Roman all the time. But everyone
at work calls him my dad, except me.
No, it's wonderful. Like, if he, if he had the right he's working with a family
member it's the best thing on the planet because you know he's not gonna fuck you
over at any point when your interests are above anything else.
That was it. The new episode of the Iliac Suite and while I'm talking the world of
AI is turning fast again. New stories on the horizon, new platform -superioring new
problems creating. If you want to stay updated, tune in this little podcast.
Thanks Ilya for talking to me. Thank you also Delphos for giving me all the music
of this episode. Thanks for listening humans. Take care and behave.