AI Amp Modeling Next Gen: A Frankenstein Scenario

[AgingEngineer]

I'm going to propose a "what if" scenario, but need to explain the backstory first. My amp rig has evolved over the last 30 years and today I'm using the Line 6 Helix for my entire signal chain for live shows, and am very happy with it.

Yet, like any gearhead with GAS, I am always looking at the next cool toy.

A thought struck me: There has been a lot of modeling amps released since the Helix, but it seems like we've hit a plateau. Pretty much any modern modeling amp will give you great tone, so anytime I listen to a new unit, there's no surprises there. So, what's next then?

As a joke --because in my real life I am a systems engineer in the IT / manufacturing space and every single corporation is spending millions scrambling to incorporate some sort of AI to outrun their competitors-- I said to a friend:

"I'm just waiting for someone to introduce an AI modeling amp. You'll train a patch on a specific player's style and tone, then the AI will real-time generate output that sounds like the artist you trained. You could be a complete novice barely pounding out a recognizable riff, and it comes out sounding perfect. On pitch, in time, AI generated riffs and fills in the right places, sounding just like that trained artist no matter how poor the input."

It was funny until it struck me that if anyone thinks anything, that means someone is probably already working on it.

 

[DirtDog]

I wonder what kind of computing power would be needed for a use case like that. And would that fit into a form factor we are accustomed to? And what would the power supply need to be?

 

[DirtDog]

It was funny until it struck me that if anyone thinks anything, that means someone is probably already working on it.

I’ve experienced this exact same thing a thousand times in my career.

 

[AgingEngineer]

I wonder what kind of computing power would be needed for a use case like that

Probably less than we think it will! For example, Stable Diffusion was trained on terabytes of images scraped from all corners of the web, but everything it learn distilled down to a model less than 4gig that easily runs on halfway decent PC.

 

[gtirard]

I think it's probably more complex than you think, because there's a question of sound (amp simulation and the like) but also a question of fixing the playing technique. In real time. For example, if you play a note ahead of time, it's easy (in theory) to delay that note to fix its timing, but what about if you play a note too late? A many other questions I haven't thought of yet.

And, ultimately, even if it was feasable, would it be desirable, or is it distopy?

 

[AgingEngineer]

@gtirard: I agree. I couldn't image the complexity needed to do that. I try to keep up with AI, machine learning, and neural network tech but it's too much and changing too fast.

Is it feasible? Right now, probably not. In 5 years, I am willing to bet good money that it is.

If you told me 10 years ago that I could give a computer a sample of my writing style and voice and it would be able to churn out "new" content and speak it in my voice in a way that was indistinguishable from my own, I'd have said no way. Yet I routine do that today. How does AI learn not only my voice, but my speaking mannerisms and even mouth noises? I have no idea. Even so, it's a readily available reality.

Is it desirable? To me: no. On one hand tools like autotune, time correction, and automated home studio production tools have enabled anyone to create music, but the flip side of that is a sort of plastic, insincere... sameness... of everything. That human originality gets supplanted by slick formulaic production, and the sense of wonder and surprise is lost.

But... if I were a kid who wanted to be onstage with a band without investing the years of practice and blisters to be a decent musician, I'd be very interested.

 

[gtirard]

@gtirard: I agree. I couldn't image the complexity needed to do that. I try to keep up with AI, machine learning, and neural network tech but it's too much and changing too fast.

Is it feasible? Right now, probably not. In 5 years, I am willing to bet good money that it is.

If you told me 10 years ago that I could give a computer a sample of my writing style and voice and it would be able to churn out "new" content and speak it in my voice in a way that was indistinguishable from my own, I'd have said no way. Yet I routine do that today. How does AI learn not only my voice, but my speaking mannerisms and even mouth noises? I have no idea. Even so, it's a readily available reality.

Is it desirable? To me: no. On one hand tools like autotune, time correction, and automated home studio production tools have enabled anyone to create music, but the flip side of that is a sort of plastic, insincere... sameness... of everything. That human originality gets supplanted by slick formulaic production, and the sense of wonder and surprise is lost.

But... if I were a kid who wanted to be onstage with a band without investing the years of practice and blisters to be a decent musician, I'd be very interested.

Yeah, but why would a kid study (or even just play) music if an AI can generate it for him?
For me, it's like if you could ask to an AI to finish a video game for you… What's the point?

 

[InhumanResource]

Probably less than we think it will! For example, Stable Diffusion was trained on terabytes of images scraped from all corners of the web, but everything it learn distilled down to a model less than 4gig that easily runs on halfway decent PC.

Wouldn’t need much if there’s a data connection, would it? Just Wi-Fi and a login to BassGPT premium, lol.

 

[Jebberz]

I can imagine this as a plug-in, like an adavanced auto-tune (probably an Auto-Tone and an Auto-perform), but in real-time without noticeable latency, i'm not sure we are quite there yet.

 

[AgingEngineer]

Just Wi-Fi and a login to BassGPT premium, lol.

And maybe "bass controller" like from the Guitar Hero games!

 

[gtirard]

I'm going to propose a "what if" scenario, but need to explain the backstory first. My amp rig has evolved over the last 30 years and today I'm using the Line 6 Helix for my entire signal chain for live shows, and am very happy with it.

Yet, like any gearhead with GAS, I am always looking at the next cool toy.

A thought struck me: There has been a lot of modeling amps released since the Helix, but it seems like we've hit a plateau. Pretty much any modern modeling amp will give you great tone, so anytime I listen to a new unit, there's no surprises there. So, what's next then?

As a joke --because in my real life I am a systems engineer in the IT / manufacturing space and every single corporation is spending millions scrambling to incorporate some sort of AI to outrun their competitors-- I said to a friend:

"I'm just waiting for someone to introduce an AI modeling amp. You'll train a patch on a specific player's style and tone, then the AI will real-time generate output that sounds like the artist you trained. You could be a complete novice barely pounding out a recognizable riff, and it comes out sounding perfect. On pitch, in time, AI generated riffs and fills in the right places, sounding just like that trained artist no matter how poor the input."

It was funny until it struck me that if anyone thinks anything, that means someone is probably already working on it.

Have you heard about NAM? Neural Amp Modeler.

It's AI applied to amp modeling. Create your own amp models from your hardware amps, share, and download other people's models. Think Helix, but open source and free.
It's only one tiny part of your "dream", but it's available now.

OK, now the AI can simulate any amp, you just have to find a way to simulate any musician

You can also play with covers.ai in the meantine.

I think NAM is more useful than covers.ai, but anyway all this is scary.

 

[Bassed in NZ]

Band-in-a-Box is an early precursor - I remember doing a project probably 15 years ago where an outro bass solo was required - put in the chord changes, select a basic vibe and tell it you want something that sounds like Jaco - and it did a pretty creditable job!
Granted, not real time, and it wasn’t processing played notes, but it was impressive to me at the time.
And not a small leap from say vocal processors that already detect chords to create harmonies to coming up wth lines on the spot: “Just play the root and I’ll take it from there” - Skynet

 

[agedhorse]

I think the bigger issue is the reason players play a bass… for the challenge and personal creative reward.

Would enough players accept the new technology, and would it provide the benefits that justify the considerable cost?

 

[-Asdfgh-]

I'm going to propose a "what if" scenario, but need to explain the backstory first. My amp rig has evolved over the last 30 years and today I'm using the Line 6 Helix for my entire signal chain for live shows, and am very happy with it.

Yet, like any gearhead with GAS, I am always looking at the next cool toy.

A thought struck me: There has been a lot of modeling amps released since the Helix, but it seems like we've hit a plateau. Pretty much any modern modeling amp will give you great tone, so anytime I listen to a new unit, there's no surprises there. So, what's next then?

As a joke --because in my real life I am a systems engineer in the IT / manufacturing space and every single corporation is spending millions scrambling to incorporate some sort of AI to outrun their competitors-- I said to a friend:

"I'm just waiting for someone to introduce an AI modeling amp. You'll train a patch on a specific player's style and tone, then the AI will real-time generate output that sounds like the artist you trained. You could be a complete novice barely pounding out a recognizable riff, and it comes out sounding perfect. On pitch, in time, AI generated riffs and fills in the right places, sounding just like that trained artist no matter how poor the input."

It was funny until it struck me that if anyone thinks anything, that means someone is probably already working on it.

We considered how to implement AI for amp modelling, if not seriously (you wouldn't have been able to have done it in real time back then) in 1997 when I was an AI researcher. We obviously didn't try implementing it. But I was working on AI techniques to model complex functions at the time, so it seemed fair game.

 

[Passinwind]

I think the bigger issue is the reason players play a bass… for the challenge and personal creative reward.

Would enough players accept the new technology, and would it provide the benefits that justify the considerable cost?

Getting the AI chicks is worth whatever it costs...

 

[AgingEngineer]

@Bassed in NZ: Oh, band in a box! I forgot about that one. Played around with it years ago on a whim, but found myself disappointed. I think I was hoping it would turn my half-baked ideas into something great, and what it churned out was better orchestrated half-baked ideas. lol

@agedhorse: I can see multiple sides to that. Seasoned players would have no use for it. Kids who want the instant win without the effort will want one just to feel better than they are. The techno-hackers... Now that is the group I would watch: they'd train the models in unexpected ways. Who knows what kind of new sounds or performance techniques might come out of non-traditional training of a tool like that?

 

[-Asdfgh-]

Probably less than we think it will! For example, Stable Diffusion was trained on terabytes of images scraped from all corners of the web, but everything it learn distilled down to a model less than 4gig that easily runs on halfway decent PC.

We just looked at something closer to what is now a Kemper. Adding style effects would be complex in terms of generating a training set. I don't know how you'd do it

 

[Marko 1]

Would enough players accept the new technology, and would it provide the benefits that justify the considerable cost?

I'll wait 'till it's used at GC for cheap.

 

[N. Kentucky Bass]

I think the bigger issue is the reason players play a bass… for the challenge and personal creative reward.

Would enough players accept the new technology, and would it provide the benefits that justify the considerable cost?

New music creation tools are now being pushed with the message, "be a music creator without having to hassle with learning music theory or how to play an instrument." I believe that message resonates with many people who are into clicking on something and getting the result in two seconds. TikTok has figured it out.

 

[ardgedee]

You could be a complete novice barely pounding out a recognizable riff, and it comes out sounding perfect. On pitch, in time, AI generated riffs and fills in the right places, sounding just like that trained artist no matter how poor the input.