NAM: Neural Amp Modeler

[Orvillain]

Another great video from Leo. He nulled the real amp against itself to show that a tube amp doesn't achieve a perfect null with the same DI, and then trained NAM up to 10k epochs to see how the ESR decreases with more epochs (it eventually hits an asymptote). He also compared that against the real amp's self-null, and calculated how long it would take to train all knob permutations if you segmented each knob to a limited set of discrete positions.

Well I learned a lot about ... erm... archery....

 

[SwirlyMaple]

As soon as we get multi-capture captures (IE: proper bass/mid/treble/presence/depth/additional switches of the actual amp) then I think capture loders will be table stakes.

And yes it's a lot of work.

But how the flip does anyone think their drum software is being made? Multiple weeks in a studio hitting drums like a dummy, and then editing it all up, that's how!!

Yeah. The biggest limitation right now is just the computational time required to crunch all those captures. But with a decent GPU array (like a crypto mining stack), that becomes less of a problem. It will still take a *while,* but when you consider Line 6 takes a month per amp model, multi-capture models could soon be in that realm. And once that exists, there isn’t much reason to model amps anymore, as a traditional model isn’t likely to ever be quite as accurate as a trained neural network model.

 

[2112]

I'll be curious to see if @polyeffects can get NAM running in Beebo. Loki's UI designs for several open source synth modules make them much easier to use. Plus, I already own a Beebo so it would be fun to have that able to run a capture of my Mark IV.

 

[the-trooper]

Yeah. The biggest limitation right now is just the computational time required to crunch all those captures. But with a decent GPU array (like a crypto mining stack), that becomes less of a problem. It will still take a *while,* but when you consider Line 6 takes a month per amp model, multi-capture models could soon be in that realm. And once that exists, there isn’t much reason to model amps anymore, as a traditional model isn’t likely to ever be quite as accurate as a trained neural network model.

Damn, didn't think about any of this. Good points.

 

[James Freeman]

traditional model isn’t likely to ever be quite as accurate as a trained neural network model.

I mostly agree but not entirely.
Captures will never capture the time variant response of the poweramp and power supply (bias excursion, power supply sag, dynamic crossover distortion, etc.).
Captures will never be as versatile and tweakable as full component models.

I predict hybrid approach at first, where the preamp is a capture, tonestack is component based, then poweramp which is also a capture.

 

[BenIfin]

So do you all think Kemper and other expensive capturing modelers will go-in-the-way-of-the-dodo?
Do you think capture loaders will be table stakes as IRs have become?

There is no stopping it, the cat is out of the bag and completely open source under MIT license.
We have yet to see a NAM profile loader pedal, but we will, sooner than we expect.

Putting the whole "in-house" -vs- "open sourced" AI modeling debate aside, IK via Tonex have already led the charge on AI based modeling.

Ben

 

[SwirlyMaple]

Captures will never capture the time variant response of the poweramp and power supply (bias excursion, power supply sag, dynamic crossover distortion, etc.).

That’s a great point! But… never say never I don’t think there’s any theoretical limitation as to why neural networks can’t be used for time-variant behavior. It just hasn’t been the research priority yet.

Captures will never be as versatile and tweakable as full component models.

Agreed to an extent here too… except that it’s probably feasible to use neural nets to represent discrete parts of an amp with higher accuracy, too.

Working forward from the physics principles to model something can get you very close, but it’s those little hair-splitting things that need many more degrees of freedom which limits traditional modeling.

If you used a neural net to model chunks of circuits based on measured input and output, and then had a way to combine those chunks, that might be another way that neural nets could approach full models but still have adjustable parameters.

 

[Orvillain]

I predict hybrid approach at first, where the preamp is a capture, tonestack is component based, then poweramp which is also a capture.

I was literally just typing that out!

 

[James Freeman]

Yeah the hybrid approach is the logical intermediate step.
But I've said before there is a point of diminishing returns with hybrid captures where is it easier to do a full component model even if it will be slightly less accurate than the actual amp captured, but it will have all the switches being schematic based.

Personally I prefer the full functionality of the amp more.

 

[Orvillain]

That Facebook group is rife with cringe people.

 

[MirrorProfiles]

As soon as we get multi-capture captures (IE: proper bass/mid/treble/presence/depth/additional switches of the actual amp) then I think capture loders will be table stakes.

Like Neural DSP and Mercuriall?

Given they’ve both had products out for a few years now, I don’t think people generally think they’re better or worse than the algorithm based alternatives.

I think in either case we hit a plateau where user error holds things back, and the reluctance of certain users to understand their tools means we don’t advance further….

 

[MirrorProfiles]

Facebook really is a terrible platform to try and suggest that some people might be doing something wrong.

I’ve thought it for a while, but I would really rather have a paid tier of a product built off NAM that has some kind of structure and purpose to it. The open source aspect is amazing and has its value, but it doesn’t really lead to it becoming a good product. It’s just good tech until someone does the work involved in making it a product.

 

[MirrorProfiles]

also one huge advantage of algorithm based sims is how you can build and tweak circuits that don’t require a real world model. When I think of the possibilities of what a digital emulation should offer, then I don't think it makes sense to always be so totally constrained by owning a physical unit that can only be captured (and constrained) under real word conditions. There even needs to be some amount of understanding of circuits - do you model the entire amp when you adjust power amp valves or bias? what about for each switch on the amp - what if they are relays that affect things in different part of the amp?

Adding gain stages, diode clipping, master volume topologies etc. Not saying it can’t be done with ML, but it would be a ton of work with its own problems and presumably not worth the effort?

I think hybrid is the way forward, on a case by case basis. Another problem is when you isolate certain blocks, you can miss important interactions between sections. Genuinely curious of ML can be trained to write more accurate code for algorithms

 

[MirrorProfiles]

I've written a small .pdf about calibration and why we should think about it.

Does anyone nerdy enough fancy checking over it, and making sure I'm not talking absolute bollocks?

or do I just dump it on @James Freeman as usual?

(HAPPY EASTER EVERYONE)

 

[BenIfin]

Facebook really is a terrible platform to try and suggest that some people might be doing something wrong.

I’ve thought it for a while, but I would really rather have a paid tier of a product built off NAM that has some kind of structure and purpose to it. The open source aspect is amazing and has its value, but it doesn’t really lead to it becoming a good product. It’s just good tech until someone does the work involved in making it a product.

^^^ In my view, you have perfectly summarized the key point about NAM and every other emerging potentially open-sourced AI modeler.

These are the very reasons why Tonex is such an awesome sounding and feeling and supported unit and eco-system.

Then think about when it was released and made available to buy and use to "the public" - in short, its still in its absolute infancy and will doubtlessly get better and better and better etc.

I know this is not a popular view from those that are all-in on the "open sourced" approach, but i.m.h.o, the only real chance that NAM-like products have to compete at a "pro" level, is for a "big-player" to adopt and incorporate and develop them into their own software and hardware products - unless/until that happens, its a free and fun tech to play with.

Ben

 

[James Freeman]

Does anyone nerdy enough fancy checking over it, and making sure I'm not talking absolute bollocks?

I'll do it.

 

[MirrorProfiles]

 

[MirrorProfiles]

any ideas on this? -3dBu seems an odd way of writing the max input level for the DI unless its relative to something?

 

[James Freeman]

Ah yes.... Behringer.

This one we'll have to measure.

 

[Orvillain]

Training a first model.