NAM: Neural Amp Modeler

[James Freeman]

NAM: Neural Amp Modeler is an open source capturing project by Steven Atkinson.

Github Main Page:

GitHub - sdatkinson/neural-amp-modeler: Neural network emulator for guitar amplifiers.

Neural network emulator for guitar amplifiers. Contribute to sdatkinson/neural-amp-modeler development by creating an account on GitHub.

github.com

VST3 Plugin:

GitHub - sdatkinson/NeuralAmpModelerPlugin: Plugin for Neural Amp Modeler

Plugin for Neural Amp Modeler. Contribute to sdatkinson/NeuralAmpModelerPlugin development by creating an account on GitHub.

github.com

Facebook Community:

Bei Facebook anmelden

Melde dich bei Facebook an, um dich mit deinen Freunden, deiner Familie und Personen, die du kennst, zu verbinden und Inhalte zu teilen.

www.facebook.com

ToneHunt:

ToneHunt | Sound Better!

ToneHunt - Find captures for NAM, Aida-X, Proteus, pedals, irs and more!

tonehunt.org

All-In-One dBu Calculator:

https://drive.google.com/file/d/1yg3FPZPMTpNwaVd1IKfuy2zKMygpk_Bl

 

[DLC86]

I'll add this too for those who have a recent Nvidia gpu and want to train locally (takes less time than google colab).

 

[BenIfin]

OK ...... in the 2nd Video " Better than Tonex " he says - he *thinks* Tonex runs something like 200 Epochs and gives and "ESR of 0.03" .... and NAM when running 500 Epochs gives and ESR of 0.017 (?)

So am I right (?) in interpreting that by saying that Tonex is %97.0 of Perfect and NAM is %98.3 of Perfect (?)

I played his demo back through my NS10's up loud and then through my Matrix GT 800 through my Celestion F12-X200 up loud ..... I could not discern or hear any difference of any kind whatsoever between the real Amp, the Tonex or Nam ?!?

I totally get numbers matter and the null tests of NAM are a *bit* better .... but even mathematically, how is anyone ever going to claim they can hear that NAM is mathematically %1.014 close to perfect than Tonex ie: %97 x %1.014 = %98.3

Am I missing something really obvious here ?

And if Tonex introduced an "Ultimate" option of [say] " 500 Epochs " ..... then who knows, maybe Tonex will be even better (?)

We are in ultra rarified air here :)

Reminds me of the "Mix Masters" Video where a Mastering engineer said that " increasing the 20khz by 0.09db was the thing that brought the mix to life" ..... I dont care how golden *anyone's* hearing is, but that always struct me as total unadulterated b.s ? :)

Yes .... but on paper ... in mathematical terms, NAM is a "fraction" better .... but i.m.h.o ..... NO-ONE will EVER be able to pick the difference in a blind test when you are operating at > %97 of "perfect"

Still interesting though.
Ben

 

[DLC86]

OK ...... in the 2nd Video " Better than Tonex " he says - he *thinks* Tonex runs something like 200 Epochs and gives and "ESR of 0.03" .... and NAM when running 500 Epochs gives and ESR of 0.017 (?)

So am I right (?) in interpreting that by saying that Tonex is %97.0 of Perfect and NAM is %98.3 of Perfect (?)

I played his demo back through my NS10's up loud and then through my Matrix GT 800 through my Celestion F12-X200 up loud ..... I could not discern or hear any difference of any kind whatsoever between the real Amp, the Tonex or Nam ?!?

I totally get numbers matter and the null tests of NAM are a *bit* better .... but even mathematically, how is anyone ever going to claim they can hear that NAM is mathematically %1.014 close to perfect than Tonex ie: %97 x %1.014 = %98.3

Am I missing something really obvious here ?

And if Tonex introduced an "Ultimate" option of [say] " 500 Epochs " ..... then who knows, maybe Tonex will be even better (?)

We are in ultra rarified air here :)

Reminds me of the "Mix Masters" Video where a Mastering engineer said that " increasing the 20khz by 0.09db was the thing that brought the mix to life" ..... I dont care how golden *anyone's* hearing is, but that always struct me as total unadulterated b.s ? :)

Yes .... but on paper ... in mathematical terms, NAM is a "fraction" better .... but i.m.h.o ..... NO-ONE will EVER be able to pick the difference in a blind test when you are operating at > %97 of "perfect"

Still interesting though.
Ben

It's not that simple. The final ESR doesn't only depend on the number of epochs but also on how you configure some parameters, on how long the test file is, on what kind of audio it contains, etc.
In my experience, regardless of number of epochs, NAM produces a more accurate result in less training time on the same hardware. And that added accuracy is definitely hearable. Tonex always has some exagerated bass frequencies and some high frequency roll-off, also it's never able to match the amount of gain automatically. NAM does everything pretty much perfectly.
I think you've seen/heard my comparison in the other thread. Well, for that model NAM took about 20 minutes on my pc to run 150 epochs with a final ESR of 0.010, Tonex took almost 45 minutes in advanced mode to produce a result with the aforementioned inaccuracies.

PS: on the other hand Tonex models are a lot lighter on the cpu than NAM models, so those inaccuracies could be some compromise IK decided to make to make it run with no issues even on older/less powerful PCs

 

[James Freeman]

I just did a capture of Helix Native using Google Colab "Easy Mode" with 100 Epochs and 0.0463 final result, it was running about 12 minutes.
Loaded into the VST and it sounded like shit.

I'm going to try 500 epochs, back in an hour.

Edit: My mistake, everything was fine, see my next post.

 

[BenIfin]

..... And that added accuracy is definitely hearable. Tonex always has some exagerated bass frequencies and some high frequency roll-off, also it's never able to match the amount of gain automatically. NAM does everything pretty much perfectly...
PS: on the other hand Tonex models are a lot lighter on the cpu than NAM models, so those inaccuracies could be some compromise IK decided to make to make it run with no issues even on older/less powerful PCs

Hey DLC86 !

I get the sense you are a big fan of NAM :) ......... %100 no issue at all with that.

In fairness ... and I fully respect these are your opinions .... I just think it is an awfully "big call" to claim the "added accuracy is definitely hearable" ..... "Tonex always has some exaggerated bass frequencies and some high frequency roll off" ...... "its never able to match the amount of gain automatically" ..... "NAM does everything pretty much perfectly".

In my opinion and to my ears, I disagree with your first point about accuracy - as I think the video I referenced above, and many other posted videos, "show" ..... I don't know of any tested A/B evidence about bass and high frequencies ...... and to my knowledge, no-one is showing / demonstrating / proving there is a "gain difference issue" - although there is a difference between the software and the pedal and that's why the pedal has a wide range input trim - this is not a problem wiht Tonex but rather because no two devices will ever be equally / identically gain-staged ....... re:- NAM being "pretty much perfect" .... not arguing that one way or the other ... these of course are all just my opinions / views.

Again, Im not in any way disputing the mathematics ....... on paper and in null tests, NAM has - what I would most generously describe as "an extremely marginal advantage".

However, I truly and genuinely believe that you or I or anyone else in a proper blind A/B testing environment, listening to, say 10 different comparison clips of each, randomly played back, would never, ever be able to pick which is which with even a remote degree of certainty .... even someone with "perfect hearing" .... is simply just not that good .,.. the human ear is awesome .... but it is just not that awesome :)

Anyway ..... I feel my comments may be construed as "Tonex useful" / "NAM pointless" .... which is actually not at all what I am trying to get at / say.

All the best,
Ben

 

[Orvillain]

Yeah but Ben, people said the same thing about the Kemper for a decade, and many of us were always able to hear the differences. I struggle to hear the difference between ToneX and NAM in the clips I've heard, but if someone else says they hear it... I'd tend to belieb them.

 

[BenIfin]

I just did a capture of Helix Native using Google Colab "Easy Mode" with 100 Epochs and 0.0463 final result, it was running about 12 minutes.
Loaded into the VST and it sounded like s**t.

I'm going to try 500 epochs, back in an hour.

Hey James !

Sorry - a bit confused ..... was the "sh$tty" Capture done with NAM or Tonex (?)

Ben

 

[BenIfin]

Yeah but Ben, people said the same thing about the Kemper for a decade, and many of us were always able to hear the differences. I struggle to hear the difference between ToneX and NAM in the clips I've heard, but if someone else says they hear it... I'd tend to belieb them.

I actually totally agree .....no two people hear the same thing the same way .... if it sounds better to "someone" then it *is* better for that "someone" regardless of why.

No doubt some people - not at all suggesting this about DLC86 for the record - reckon Tonex sounds like crap - and to them it doubtlessly does sound like crap - again the "why" is irrelevant.

I guess my only real two points are .... so far at least .....

(a) there is a growing mountain of quality reviews and clips - to my ears at least - clearly demonstrating just how amazing Tonex is ... and

(b) even though NAM *is* marginally better in terms of the math's and null tests ..... in my view ... at this level of accuracy for each "product", there is simply no way anyone - and I mean anyone - could possibly ever even remotely "pass" any sort of properly done A/B test.

All the best,
Ben

 

[MirrorProfiles]

I think the point of NAM being open source is that it’s able to continually refine and improve and minimise any differences. It’s always improving and being refined and people are able to try different ways of getting better and better results.

Whether all of these differences are always noticeable to everyone isn’t relevant - the point is that it’s advancing technology further forward with nothing to slow it down.

 

[BenIfin]

I think the point of NAM being open source is that it’s able to continually refine and improve and minimise any differences. It’s always improving and being refined and people are able to try different ways of getting better and better results.

Whether all of these differences are always noticeable to everyone isn’t relevant - the point is that it’s advancing technology further forward with nothing to slow it down.

Very true ..... I also think that the IK tech's who built Tonex are also going to continually refine and improve it to.

And if I could attempt some hopefully humorous sarcasm for a moment ..... this "open source" argument is also the exact same argument Linux users were saying and doing 15 years ago when they were all saying Linux would destroy Windows in a matter of no time due to its massive open-source user developer base .... I'd imagine the NAM open-source-user-base is a bit smaller than Linux's

All the best,
Ben

 

[James Freeman]

Okay I'm an idiot, the first capture with 100 epochs was okay, my DAW was set to 96kHz so the plugin did not sound right as mentioned on github.
The VST plays correctly only with 48kHz sampling rate at the moment.

The 500 epoch training took 58 minutes and stopped at an ESR of 0.026, it sounds nigh on identical to the source to my ear.

Capture Sharing Test:
Google Drive Download Link

Edit: The shared capture is of PV Vitriol Lead tuned for brutal chug with pre and post EQ.

 

[imperator]

From a brief look at the repo, NAM seems to be using standard model architectures like wavenet. It would be fun to see how small of a model architecture I could make and get good results.

 

[Orvillain]

Gotta love dat matplotlib

 

[DLC86]

In fairness ... and I fully respect these are your opinions .... I just think it is an awfully "big call" to claim the "added accuracy is definitely hearable" ..... "Tonex always has some exaggerated bass frequencies and some high frequency roll off" ...... "its never able to match the amount of gain automatically" ..... "NAM does everything pretty much perfectly".

It is for me and quite evident too when you directly compare it against the fm9 model, but I get not everyone would be able to hear those differences, no issue with that either.
But, just to prove it's not just my opinion, here's some facts, RTA measurements of the clips I posted:

FM9

NAM

TONEX

Look at the area around 70-80 Hz, Tonex has peaks there 1-2 dB louder than the other two.
Also look at the peaks above 6-7 KHz, tonex is 1-2 dB quieter there.

Furthermore, regarding the gain discrepancy, here's how I had to set the gain on Tonex to match the FM9 and NAM, after setting the gain trim at max (the one where you adjust the gain right after the training ends, before saving the model):

NAM doesn't have a gain adjustment before saving the model cuz it nails it with no issues, Tonex doesn't instead, it's always lower with he default mid-way gain trim setting (I'm not the only user finding that, you can see and hear the same on several youtube reviews) while after maxing it out, sometimes it's right, sometimes it's lower and other times it's higher. And I'm pretty sure it's not a routing issue cuz with the fm9 it's pretty easy to achieve unity gain entering the amp block, a further proof it was right is that the NAM capture is spot on using the same routing.

Both this inaccuracies are repeatable too, all captures I've made until now (which are not a lot btw, just 6) show the same behaviour, so definitely not just an opinion.

PS: and in case you're wondering, yes, I'm running the latest version of Tonex.

 

[DLC86]

Just to be clear, saying these things is not the equivalent to bashing tonex and IK multimedia, I'm still pretty impressed by what they're able to achieve, especially compared to older profilers/capture devices, it's just that I'm even more impressed by nam and tonex definitely has room for improvements, considering they probably run the same base code too.

PS: and just for the record, I looked at RTA graphs for the first time now so I don't think my brain is tricking me into "making me hear what I see", previously I spotted those differences just by ear and actually I expected to find bigger differences honestly.

 

[DLC86]

Okay I'm an idiot, the first capture with 100 epochs was okay, my DAW was set to 96kHz so the plugin did not sound right as mentioned on github.
The VST plays correctly only with 48kHz sampling rate at the moment.

The 500 epoch training took 58 minutes and stopped at an ESR of 0.026, it sounds nigh on identical to the source to my ear.

View attachment 4727


Capture Sharing Test:
Google Drive Download Link

Edit: The shared capture is of PV Vitriol Lead tuned for brutal chug with pre and post EQ.

when recording the output file make sure to do it at the highest possible level without clipping, that lowers the ESR even without increasing the number of epochs

 

[James Freeman]

when recording the output file make sure to do it at the highest possible level without clipping, that lowers the ESR even without increasing the number of epochs

Okay just tried that, I normalized before export.
With 100 epochs the ESR is 0.047, no improvement in ESR when it was -6db quieter, I also don't hear much difference between this and the 500 epoch capture except in output volume.
BTW this is a very high gain tone so the ESR will probably never go as low as 0.01.

 

[DLC86]

Okay just tried that, I normalized before export.
With 100 epochs the ESR is 0.047, no improvement in ESR when it was -6db quieter, I also don't hear much difference between this and the 500 epoch capture except in output volume.
BTW this is a very high gain tone so the ESR will probably never go as low as 0.01.

Normalizing is not the same as recording a hotter signal cuz by normalizing you also turn up floor and quantization noise and you gain nothing. It's expected to be the same.

PS: anyway don't expect to halve the ESR by increasing the recording level, it gets better but not by that much. If you're already happy with the results just leave it as it is.

 

[James Freeman]

Oh you mean when capturing a real amp, yeah I understand analog-to-digital gain staging well.
I was capturing Helix Native for testing which is completely in the digital domain in 32bit floating, quantization noise is irrelevant there.

Once my new DI Box arrives I will do some captures of my amp.
Prepare for balls.