So my new album "Multi-tasking" is getting ready to drop on the streaming services, and it's already on Bandcamp if you're interested. Now, let's talk about the processes that created it, and why I think it's a watershed moment, not just for me as an artist, but for AI music in general.
First and foremost, it's an album packed with AI performances. But they're different this time, relying on a number of experimental processes that were developed specifically to do this kind of music.
Of note here is that around 80% of the album was created with a process I developed called "noise painting."
I first started developing noise painting in 2019 as a way to get better selfies with AI based photo filter apps like FaceApp.
It started with the observation that facial-recognition AI will often see faces in places where a human wouldn't. So I decided to see where that line was, and it led me down this incredible rabbit hole of AI mechanics, arguing with model developers about the way nodes work, and lots and lots of really interesting visual art, as I pushed the process as far as it could possibly go.
By the end of it, I was using FaceApp to render sketches as studio photos, and I got a level of flexibility out of it that I never expected.
From FaceApp to diffusion
Then, Stable Diffusion hit. And I further developed and retooled the process in ways aimed specifically at distorting and/or manipulating the denoising mechanism in the diffusion process. I still use noise painting to make interesting visual art.
"But Lynn," you might say, "What does this have to do with music or the latest album?"
Chill, it's a long post, we're getting there.
From images to music
Chirp and Suno hit, and hard last year. It all started with me playing with the tool as a cool new AI toy. But quietly, it reinvigorated my fascination with making music in a way that was both intensely personal and profound. It completely changed my focus as an artist, when I realized how fantastically satisfying writing songs and music production can be.
I started by trying to over-prompt the system, find places nobody knew about yet, and by prompting interesting and unexpected style combinations to produce interesting results. Then, when I noticed the machine would hallucinate at weird times, I tried to figure out why. And that led me to jazz and hip hop style lyricism, and learning how to do it.
So I started to give myself these writing exercises to hone my craft, and hopefully get better at songwriting. Whether I did or not, well, that's subjective, and not for me to decide. But I have come a long way.
When Udio dropped, I was initially unimpressed. But, it quickly became a very advanced tool, and led to a lot of incredibly interesting experiments.
And there, I made a lot of perfectly adequate music, and taught myself the basics of music production the old fashioned way. I got a DAW, and a Roli keyboard, and a million bazillion plugins to do all kinds of things with sound.
I started going hard into musical compositions with Hookpad, and really started developing as an actual musician. And I started developing new processes to layer in and experiment with both AI and conventional music by layering it all in my DAW.
Noise painting with sound
The process changed when Suno and Udio, gave me file upload capabilities. Suno let me expand and extrapolate on my samples in interesting ways, and I started intentionally derailing my sound inputs to make the machine go sideways in new ways, and produce more interesting music.
Udio, however, did something interesting. They exposed so much functionality, control over the context window (not to be confused with the udio render window), and elements that screamed, “Hey, I’m a diffusion process! Use me like a diffusion process!”
And, I mean, if the app is screaming, “Hey I’m a diffusion process!” You know, there’s only one thing to do there. Yes, noise paint!
I discovered quickly that the difference between Udio’s system and Stable Diffusion are that while Stable Diffusion uses pixel noise to achieve an outcome, Udio seems to use spectral noise the same way. Which means that you can feed it a pattern of noise it’s not expecting or trained to handle, and get it to bend around sound exactly the same way you can bend Stable Diffusion with images it’s not expecting.
Noise painting with audible noise was also something I never expected to happen when all of this started, either. But that’s the big joy of working with AI. Ending up in places you never quite expected to be, and wondering blankly how you got there creatively.
I embrace this.
So I did a lot of experimenting with different kinds of sounds and recordings made in different ways. Everything from finished DAW tracks to me singing in the shower, and everything in between.
Multi-tasking
The result of these experiments, is "Multi-Tasking."
My second, or possibly third very serious, serious album.
As far as I know, nobody's done one like this yet. And I feel that I've created something rich, and natural sounding with this release.
Feel free to give it a listen.