AI and music

[willemb]

The number of Artificial Intelligence tools is increasing at a fast rate. You might have recently heard about Chat GPT and how it can write homework, code, and letters for you... and much more.

In the audio world AI is also becoming increasingly more readily available. We all can laugh at "mastering services", which might just be some ozone style processing algorhythms anyway, but there are so many other things now.

Im not approaching this from the "oh no its taking away my job" angle, as i dont make money from music. I also think AI might replace some mediocre and generic music, but human love for music will always be of value. I do think that music production and industry is going to drastically change in the coming years.

The scary factor of AI and music has already been discussed around 4.5 years ago:

http://www.subsekt.com/viewtopic.php?p=139523#p139523

Some examples of what is already available now below. Some of this might not be truly AI, some of it might be, i dunno.

Do you know of any cool or interesting examples that i missed?

https://soniccharge.com/patternarium

There is Pattenarium , with AI generated beats, a voting system, and derivatives from previously generated beats. Free to listen, you need MicroTonic to use the sounds. Its basically Microtonic patches, and you can copy the patch immediately into MicroTonic. Look through 25 beats, pick your favourite kicks, (then give it a quick edit/tweak?) rather than spending a lot more time tweaking knobs in icrotonic. I guess its not much different from having an infinite-ish list of presets, or a tweakable randomisation function... but still, check out the beats it makes, there are some pretty cool ones:

https://www.lalal.ai/

Free to listen, paid to use. It separates stems from a track. You upload a track, select Drums, and it will extract the drums for you... A way to sample things that couldnt be sampled before? Get the instrumental or acapella that you always wanted but never existed?

Atlas 2 or XO

Sample organising/finding software that group samples together to help you find sounds quicker. Im tempted to buy Atlas 2 for finding and browsing my samples quicker.

https://aiva.ai/

Composition generator? Create an account to try for free, or premium options for copy right ownership. From the examples on youtube, it creates kind of TV / cheap film background music. This might already be better for some short film or low budget video producers who just want some music to put in the background of their stuff - could be quicker and cheaper than hiring someone to make a sound track.

https://magenta.tensorflow.org/ and https://magenta.tensorflow.org/ddsp-vst

Google AI sound experiments, with free ableton thingies, vst or standalone. Generate sequences, strange synthesis and more. Just downloaded the DDSP-VST and does some stuff i havent really heard from other plugins, kind of like physical modelling type sounds, but not.

Orb Composer / Orb Producer

Idea generator (music generator), paid software. Use AI to generate chords, melodies, etc

[Amøbe]

I have recently embraced generative tools/ai a bit more. I think what they are doing with microtonic is a brilliant example on how you can make it useful and not just a gadget, which you can giggle about.

My former workplace used ai to generate playlists based on "moods". One of the issues we battled was that while it at first seems impressive, it's actually really difficult to design AI in a way that make you want to use it. Our experiments showed that the slightest doubts and misunderstandings would make most users not wanting to use it,, because they quickly assumed it was just pure randomness

[willemb]

I get the bit about not trusting it and finding it too random. I was watching reviews and demos of Atlas 2, and sometimes it would get a sound massively in the wrong category. The sample organisers available seem particular good at detecting drum basics so far: kicks, hats, shaker, snare, outside the quality isnt as good. And the (perhaps only occassional) wrong result immediately makes me doubt if i want software like that at all.



I just found this channel on youtube that is 24/7 generated riffs and music in the style of funk. Its not perfect yet, but give it a few years, and there will be endless supply of any music a computer can imagine.

NO SOUL (ai funk)

youtu.be/qGjsGTvwL78

"This is a neural network freely improvising 100% synthesized funk. It dreams basslines, choirs, changes, buildups, glitches, and surprises. At each moment, imagining the next, and so on. But it doesn't stay anywhere for long. Look around. The world is here to be played with. Pick a dream and go.

We don't know where it's going next and it never loops the same riff twice. It's evolving, unpredictable, yet I'm constantly headbanging and at times full on stankfacing. With soft gentle parts between."


I do love random buttons on plugins. If randomisers can be more specific, more tailored, im all for it. Maybe one of the most immediate, fulfilling ways we will be finding or creating sounds will soon be entirely text prompt based...

Input: Kick 128 bpm, medium length, no harsh or high click, creative pro level eq cuts, fat compression

Generates 10 kicks, pick the best ones, tell it to create variations / new generation based on those, find what you like, boom - you got a super fat kick. Future job title: Input Prompt Engineer

[nocernoc]

a lot of "ai" out there is just people pretending to be "real ai". its a marketing gimmick.

if it is really good ai, as in it's indistinguishable from a human, then it probably is human.

btw, was in the Netherlands not so long ago. drove a scooter from London to Rotterdam; cool futuristic city.
i'll go live there if UK ever gets back into the eu.
can speak basic Dutch now.
still have difficulty understanding fluent speakers though.

[willemb]

a lot of "ai" out there is just people pretending to be "real ai". its a marketing gimmick.

if it is really good ai, as in it's indistinguishable from a human, then it probably is human.

I agree, but dont be fooled that this area isnt evolving at a rapid rate. The power of Chat GPT and Dall-E is coming to audio more and more. That Funk live stream, clearly is AI, it is weird and spooky, but its getting close. One of the main issues with the audio equivalent is the size of the data. Its easy to find text and images that arent very big file size wise to train neural networks or AI with, it costs a lot more storage, memory and processing power to do the same with audio.

Ive been exploring various "AI" music things (true or not) online, some of them basically seem to be midi generators with some effects, sounds impressive when you listen to a couple of songs, but listen more than 10 and you'll hear the limitations pretty quick. Yoiu will start to recognise the sounds in the general midi kit that they use.

Things where it generates audio without midi are more spooky and impressive. (sometimes just wrong and weird, other times touching on the uncanny valley)

btw, was in the Netherlands not so long ago. drove a scooter from London to Rotterdam; cool futuristic city.
i'll go live there if UK ever gets back into the eu.
can speak basic Dutch now.
still have difficulty understanding fluent speakers though.

Netherlands (or anywhere for that matter) has its pros and cons - i do miss living in the UK at least once a week, sometimes daily. Its weird.

Basic Dutch is impressive! Its easy to just get by here with English, which makes a real hurdle to learning the language. I work exclusively with UK colleagues and customers, so i only speak Dutch in my leisure time, and my Dutch is now broken as hell. I cant remember any strong verbs for example (like i cant usually think of the equivalent of swum, so i would say swimmed, but then in dutch).

[willemb]

youtu.be/2nzSQ3up1kw

another 24/7 generated live stream from the same channel as that NO SOUL stream - mostly weird, but also cool.

They also have some Djent/Metal streams, which are kind of less impressive to me, as its seems to be more composition based and using the right samples.

[Amøbe]

a lot of "ai" out there is just people pretending to be "real ai". its a marketing gimmick.

if it is really good ai, as in it's indistinguishable from a human, then it probably is human.

This is a misunderstanding of AI. You're basically expecting AI to be human intelligence, which is also why some prefer the term "Alternative Intelligence" over Artificial.

I don't think we'll get any artificial version of human intelligence - at least that's not what they're trying to do. Instead we are getting automated systems that has an agency. This brings us to a difference between the agency of machines and of humans. At its core humans are extremely good at abstract thinking and machines are extremely fast at systematic thinking. An example of this - a few years ago everybody expected us to have self-driving car any minute now. This isn't really something we even discuss anymore, because Sweden came by and destroyed that idea. Not because of strict law making, but because of weather.

The very impressive results we have seen of self-driving cars have been done in a very dry climate. This means that the scanning equipment easily can get an overview of the road, and since roads look similar, it can look up a previous road (systematic thinking) and repeat what it knows worked the last time. Now the issue with Sweden is that it tends to snow every now and then. And when it snows, sometimes you are not able to see the road - and then you can't look it up in the database and see how to proceed. Now humans are able to drive when it snows (yes we have to drive slower, and sometimes we do in fact crash, but an impressive amount of time we don't). What the human brain is capable of doing is to approximate where the road probably is. Having to do calculations like this in real time (not looking it up) is insanely demanding for a machine - and right now it can't do that.

So humans are impressively fast at meeting a problem we haven't dealt with before, and then figuring out how to proceed. Machines are insanely fast at doing tasks in a big system (give me the most expensive building lots in Britain, give me the most promising stocks in the world etc.), where it can work based on previous actions.

Now some might say that's not intelligence, but honestly that's a very anthropocentric view of intelligence. The universe figured out how to exist without any humans interacting, genomes figured out which mutations to keep breeding in order to adapt to changes and so on. A - at least if you want to dive into the field - more fruitful question is not "how is it intelligent like me (a thinking monkey)", but "what agency can it perform on its own".

Then you're also in a better position to critique it, because as with all things there are dangers. Most AIs we use today has a tendency to become super racist really fast. For instance at one point I worked on a feature where music should be recommended based on imagery. It had been trained (amongst other things) on music videos, and then my boss kept being annoyed with it, because it kept recommending him hip hop (which he wasn't really into). At first we couldn't figure out why that happened, because it was not an issue to the rest of us, until someone by coincidence discovered that it connected black people with hip hop (due to the music videos), and my boss had dark skin - so we had to train it to be a bit more diverse

The thing is - we didn't try to give it that bias - there was just a strong systematic connection in the ground truth it had been trained upon, and there are loads of issues like this, where - even though you try to prevent it by checking the data beforehand - the AI sees some connection between some data and ethnicities... which is a good way of discovering systematic biases in our society and culture in general, but that's an even longer discussion.

Point is - humans don't think like computers do. What we call artifical intelligence could as well be called very-complex-automations, but you risk the danger of shutting down the entire discussion of AI, if you only recognize it as how we humans interact with life.

[Amøbe]

youtu.be/2nzSQ3up1kw

another 24/7 generated live stream from the same channel as that NO SOUL stream - mostly weird, but also cool.

They also have some Djent/Metal streams, which are kind of less impressive to me, as its seems to be more composition based and using the right samples.

Just listened a couple of minutes - that's amazing it sounds kinda like something that misunderstood music trying to make music (which is what I really want from AI if it should generate on its own)

[willemb]

The variety of sounds/styles it comes up with is quite amazing, you should check back at some different times. It sometimes is really just weird noise, other times it actually sounds like real songs.

[chava]

I am quite surpised that "AI" or algorithmic generated music hasn't gotten farther than it is. Maybe the data issue is key as noted above. What I still from examples is that this could go fine as mediocre background music in tv/film, games, cafes and crappy clubs. But it is not good in even doing proper muzak yet! So still waiting, but not really in excitement I must say.

[Amøbe]

I am quite surpised that "AI" or algorithmic generated music hasn't gotten farther than it is. Maybe the data issue is key as noted above. What I still from examples is that this could go fine as mediocre background music in tv/film, games, cafes and crappy clubs. But it is not good in even doing proper muzak yet! So still waiting, but not really in excitement I must say.

I think there's two main reasons for that (which I'm admittedly pulling a bit out of my ass):

1) it being based on music that already exists makes it kinda hard to get to do something that's actually new and still musically meaningful, and not just aping other styles. (can you use ape as in copy in English?)

2) there's always more than the music itself going on. Whether that's the context we consume it in (car, church, club, forum-recommendation etcetera), or the myths we are creating on the artist (i.e. a good example of this is why some of us still wants to hear everything autechre makes. There are other max convoluted music out there, but there's something about following these two blokes and admire their musical blossoming or something. Another example is the pop-/rockstar)

AI music - what I've heard - is 99 % pure music. It sort of exists without a purpose or any connection to the society I'm living in. And I need that for music to be elevated to art like Britney Spears, who I think for 20 years was the best vessel of society to express itself through.

[willemb]

TV/youtube/cheap game or film can be done by background composition generator sites/software, like:

https://mubert.com/render

https://score.ampermusic.com/

https://ecrettmusic.com/play

Sites like this basically generate midi and song structure, maybe throw on a couple of simple effects. I was looking at all of em a few days ago, i cant remember which one, but there was one that really looked worth it with a huge library of generated songs that would suit many basic needs.

The feeding of examples is more of an issue when working with audio itself, rather than with midi/song structure. I am busy getting hooked into the world of AI, and looking into making models to generate my own samples... feed 2000 kicks into a model and let it spit out diffused recreated kicks, etc.

[Amøbe]

TV/youtube/cheap game or film can be done by background composition generator sites/software, like:

https://mubert.com/render

https://score.ampermusic.com/

https://ecrettmusic.com/play

Sites like this basically generate midi and song structure, maybe throw on a couple of simple effects. I was looking at all of em a few days ago, i cant remember which one, but there was one that really looked worth it with a huge library of generated songs that would suit many basic needs.

The feeding of examples is more of an issue when working with audio itself, rather than with midi/song structure. I am busy getting hooked into the world of AI, and looking into making models to generate my own samples... feed 2000 kicks into a model and let it spit out diffused recreated kicks, etc.

But that's also using it as a tool, where you collaborate with it, rather than it is supposed to be its own creative entity - I find the collaborative aspect much more interesting in general, because I can definitely imagine situations, where that would make sense

[terryfalafel]

still have difficulty understanding fluent speakers though.

I've been here for ten years and spent thousands on language courses and still have the same problem

[Ben Kohonays]

(can you use ape as in copy in English?)

Yes!

[Amøbe]

(can you use ape as in copy in English?)

Yes!

Weehoo!

[chava]

I am quite surpised that "AI" or algorithmic generated music hasn't gotten farther than it is. Maybe the data issue is key as noted above. What I still from examples is that this could go fine as mediocre background music in tv/film, games, cafes and crappy clubs. But it is not good in even doing proper muzak yet! So still waiting, but not really in excitement I must say.

I think there's two main reasons for that (which I'm admittedly pulling a bit out of my ass):

1) it being based on music that already exists makes it kinda hard to get to do something that's actually new and still musically meaningful, and not just aping other styles. (can you use ape as in copy in English?)

2) there's always more than the music itself going on. Whether that's the context we consume it in (car, church, club, forum-recommendation etcetera), or the myths we are creating on the artist (i.e. a good example of this is why some of us still wants to hear everything autechre makes. There are other max convoluted music out there, but there's something about following these two blokes and admire their musical blossoming or something. Another example is the pop-/rockstar)

AI music - what I've heard - is 99 % pure music. It sort of exists without a purpose or any connection to the society I'm living in. And I need that for music to be elevated to art like Britney Spears, who I think for 20 years was the best vessel of society to express itself through.

ad 1) But everything "original" (let's call it human generated) is also in some way aping other styles. No one is truly original, so theoretically the AI system should be able to do something as ingenious as a human being.

.. 2) should not really be relevant for the dj scene, or should it?

[willemb]

About copying, the weird thing is that the patterns it finds arent always the patterns we can find, or we can even describe. It appears that diffusion is good at getting timbre or expression or a sound of a style... Train a modle with the beatles, and it might spit out something that is totally incoherent, doesnt sound like a song, doesnt have proper melodies, no recognisable words, but you can hear its the beatles. I guess it sounds like lots of beatles samples that are mangled, and i guess that is what it is, but its mangled in a special AI way.

https://dadabots.bandcamp.com/track/evo ... tion-60501

[chava]

MusicLM - New AI music project from Google:

https://google-research.github.io/seane ... /examples/

[Lost to the Void]

I am quite surpised that "AI" or algorithmic generated music hasn't gotten farther than it is. Maybe the data issue is key as noted above. What I still from examples is that this could go fine as mediocre background music in tv/film, games, cafes and crappy clubs. But it is not good in even doing proper muzak yet! So still waiting, but not really in excitement I must say.

I think there's two main reasons for that (which I'm admittedly pulling a bit out of my ass):

1) it being based on music that already exists makes it kinda hard to get to do something that's actually new and still musically meaningful, and not just aping other styles. (can you use ape as in copy in English?)

2) there's always more than the music itself going on. Whether that's the context we consume it in (car, church, club, forum-recommendation etcetera), or the myths we are creating on the artist (i.e. a good example of this is why some of us still wants to hear everything autechre makes. There are other max convoluted music out there, but there's something about following these two blokes and admire their musical blossoming or something. Another example is the pop-/rockstar)

AI music - what I've heard - is 99 % pure music. It sort of exists without a purpose or any connection to the society I'm living in. And I need that for music to be elevated to art like Britney Spears, who I think for 20 years was the best vessel of society to express itself through.

ad 1) But everything "original" (let's call it human generated) is also in some way aping other styles. No one is truly original, so theoretically the AI system should be able to do something as ingenious as a human being.

.. 2) should not really be relevant for the dj scene, or should it?

Nah, people who are into music can tell the difference between production line pop, made by committee and something with general heart and depth.
AI might be great for pop artists. Cheaper than paying writers and producers. You just need a pretty face, a functional singing voice, charisma and marketing/doing viral shit, and an AI providing the backing music.

Not sure AI is going to fill many stadiums.