FM9 vs NAM vs TONEX comparison

DLC86

Roadie
Messages
324
Hi all, just discovered this place and I'm glad we now have a valid alternative to tgp to discuss gear.

Yesterday I made a little comparison between the aforementioned products that I already posted on Fractal forum, but this seems a better place to discuss this sort of things.
Basically I recorded a DI track with some chugging and guitar volume roll-offs, then reamped it thru the FM9 VH4 model and captures I made of it on both NAM and ToneX. Here are the results:

FM9


NAM


TONEX


FM9-NAM null test


FM9-TONEX null test


It seems both NAM and ToneX can nail the dynamic behaviour of the FM9, but while NAM nails everything pretty much perfectly, ToneX still sounds a bit different especially on the low and high ends.
Also, it's been a nightmare to precisely match the gain amount on ToneX, for some reason it's not able to match it by itself while NAM does it with no effort.
The accuracy of NAM is quite evident even looking at the waveforms, it looks almost identical to the FM9 clip while ToneX shows a lot of differences.

FM9-NAM-TONEX.jpg



I'm quite impressed by NAM, it produces an almost indistinguishable result in less training time than the advanced mode of Tonex. What do you guys think? Have you tried it yet?
 
NAM is really great, certainly more accurate than any other method I'm aware of for reproducing tones. The speed of development and community is also really good at the moment. Steve's attention to detail and self critique to get the accuracy as good as he possibly can is awesome and something I'm totally behind. He's also so attentive and quick to try and please everyone, I have no idea how he manages.

My main issues with it are really just from the nature of it being open source and constantly in development - the goalposts are always moving and there's always a bit more fudging around than you'd get from a commercial product. Its one of those things where its strongest attribute is also its biggest flaw - I'd actually prefer if it was a commercial product so it came with the assurances of stability, support and usability that a commercial product would have.

Big fan of what Steve is doing though and I'm always following along with the developments. Much in the way that Kemper's still offer something valid, the slight trade off in accuracy from ToneX is kind of balanced out by the fact its packaged up in a way that is a bit more dependable.
 
I hope he'll be teaming up with some hardware company - without ever losing his independence.
Anything I've heard from NAM so far is pretty much fantastic. Too bad I can't run the latest version(s) on my machine (an aging Mac Pro).
 
The interesting thing about NAM is that it's heavily leveraging the PyTorch library and the NAM developer is mainly fine tuning how it is configured - like majority of the code is to feed the data to the library and define how that data is weighted. Now I'm not saying that is not a significant amount of work, but it is a very different paradigm from component modeling.

I could never do what e.g Cliff Chase does in component modeling, but with some learning I could probably make my own NAM if I put in the time and effort because I don't need to know about the intricacies of amp circuits and how to reproduce them in the digital realm.

For NAM, most of the harder work is on the actual plugin side to make it do more than just run the model and an IR.

I can see eventually these sort of models becoming a building block on the cheaper amp modelers like Nux etc, leapfrogging how accurate they are. If atm e.g the Nux Amp Academy is not quite as good as the big boys in digital modeling, the next iteration could very well be quite close. I don't mean they are Tonex type profile loaders, but instead offer pre-built models you choose from.

What that means for the more expensive products from Line6 or Fractal will be interesting. Obviously both companies have the advantage of "it also operates like the real amp" vs these "just a snapshot" models, but I think many don't care about that distinction so much as long as they get the tones they like and can shape them to their liking. Another benefit to the expensive boxes is all the fx they can do and the control options for managing them together which is more work on a pedalboard setup.
 
Spoilt for choice these days!! NAM sounds interesting, I've not tried it yet, but I'll have to set aside an afternoon to play around with it. I'm familiar with Python and PyTorch myself, so doesn't look too hard.
 
The interesting thing about NAM is that it's heavily leveraging the PyTorch library and the NAM developer is mainly fine tuning how it is configured - like majority of the code is to feed the data to the library and define how that data is weighted. Now I'm not saying that is not a significant amount of work, but it is a very different paradigm from component modeling.

FWIW, this is also how ToneX works. If your check your installation you'll find a pretty standard PyTorch setup under the hood.
 
Hi all, just discovered this place and I'm glad we now have a valid alternative to tgp to discuss gear.

Yesterday I made a little comparison between the aforementioned products that I already posted on Fractal forum, but this seems a better place to discuss this sort of things.
Basically I recorded a DI track with some chugging and guitar volume roll-offs, then reamped it thru the FM9 VH4 model and captures I made of it on both NAM and ToneX. Here are the results:

FM9


NAM


TONEX


FM9-NAM null test


FM9-TONEX null test


It seems both NAM and ToneX can nail the dynamic behaviour of the FM9, but while NAM nails everything pretty much perfectly, ToneX still sounds a bit different especially on the low and high ends.
Also, it's been a nightmare to precisely match the gain amount on ToneX, for some reason it's not able to match it by itself while NAM does it with no effort.
The accuracy of NAM is quite evident even looking at the waveforms, it looks almost identical to the FM9 clip while ToneX shows a lot of differences.

View attachment 4699


I'm quite impressed by NAM, it produces an almost indistinguishable result in less training time than the advanced mode of Tonex. What do you guys think? Have you tried it yet?


Hey DLC86 !

Welcome to TGF :)

I agree. NAM is amazing. I have messed with it too and concur with your results / conclusions :)

Some "side" comments ......

Of the currently available products to use, Tonex is the clear leader - see here for a Real Amp vs QC Capture vs Tonex Capture each Null tested against the real amp and each other:- https://youtu.be/Avhih6805YU?t=1157 - the KPA is a long way off on the null test compared to the real amp .... the QC is a bit better ..... but the Tonex is in a league of its own compared to the real amp and the KPA and QC.

When I bought my KPA, I clearly remember many posts from Christoph on the KPA private forum about people "complaining" that their real amp profiles were excellent but when they tried to profile another modeler or software, the results were much worse - Christoph repeatedly confirmed this to be true and explained / stated many times that *all* his design and testing for the KPA was done with real amps and that it definitely *would not* properly capture other modelers or software emulations ..... I don't know for sure, but I strongly suspect this is exactly the same situation with the Tonex environment (?)

Guitarjon [ SDS ] recently posted a Y/T video where he Tonex Captured his JCM 800 ... and the results were quite literally indistinguishable ... he stated that even he could not hear or tell the difference

Is NAM "the best" via null tests now ..... yep !

Is Tonex noticeably better than the KPA or QC .... yep !

Even allowing for all the above, is Tonex so close / identical enough that it doesn't matter .... yep !

I.m.h.o ...... the real-world "difficulty" for NAM is that without a "major backer" and "implementer" taking them onboard, the NAM "process" probably ain't going to get any traction anywhere ..... even if a "major backer" and "implementer" took it on now, we are looking at at least 2+ <-> 3+ minimum years before we *could maybe* see *some sort of hardware* that was live-useable ..... and in the current environment, and the niche MI marker, of which digital modeling is an even much smaller niche .... I just cant see NAM going anywhere in terms of functional useability for any level for musicians other than strictly "home-based-nerds" <= no offence meant to anyone.

And even then, Tonex wont stand still ... it will keep getting better .... they have already improved the algorithm for High Gain Capturing in the first week of its release ....... the QC *may* lift their overall game ... and who on earth has any idea what Christoph is going to do ? :)

So i.m.h.o .... if anyone wants the best possible Capture / Profiling Tech to use live, in the studio or on tours etc.... ... its the Tonex ..... and I think, notwithstanding the above, this is going to be the case for many years to come.

But yes ..... NAM is amazing ! :)

All the best,
Ben

PS: I.m.h.o ..... the only logical practical place for NAM is to be somehow licensed by Line 6 for inclusion into the Helix or next top tier modeler [not Helix 2 ] that they are doubtlessly working.
 
Last edited:
FWIW, this is also how ToneX works. If your check your installation you'll find a pretty standard PyTorch setup under the hood.
Interesting. Since both are using the same library, it mainly comes down to the developer of NAM having done a better job at configuring it to his bidding. This on the other hand means that IK can also improve Tonex over time, even though it means having to redo captures.
 
The interesting thing about NAM is that it's heavily leveraging the PyTorch library and the NAM developer is mainly fine tuning how it is configured - like majority of the code is to feed the data to the library and define how that data is weighted. Now I'm not saying that is not a significant amount of work, but it is a very different paradigm from component modeling.

I could never do what e.g Cliff Chase does in component modeling, but with some learning I could probably make my own NAM if I put in the time and effort because I don't need to know about the intricacies of amp circuits and how to reproduce them in the digital realm.

For NAM, most of the harder work is on the actual plugin side to make it do more than just run the model and an IR.

I can see eventually these sort of models becoming a building block on the cheaper amp modelers like Nux etc, leapfrogging how accurate they are. If atm e.g the Nux Amp Academy is not quite as good as the big boys in digital modeling, the next iteration could very well be quite close. I don't mean they are Tonex type profile loaders, but instead offer pre-built models you choose from.

What that means for the more expensive products from Line6 or Fractal will be interesting. Obviously both companies have the advantage of "it also operates like the real amp" vs these "just a snapshot" models, but I think many don't care about that distinction so much as long as they get the tones they like and can shape them to their liking. Another benefit to the expensive boxes is all the fx they can do and the control options for managing them together which is more work on a pedalboard setup.
Yeah, I agree that this kind of captures based on neural networks are kinda easier to program than a full white box simulation, and we might see a lot other products in this category in the near future. But at the same time it seems it's not that easy to configure it properly either since neither Tonex nor QC are able to produce such accurate results, at the moment at least.

For me component modeling still has a place, with that approach you can sculpt the tones however you want vs being limited to use "snapshots", and that's why I'll never abandon Fractal (in addition to FX and routing options), but imagine what a neural network applied to an accurate base "white-box" model could achieve: let the user choose a model that is the equivalent or the closest to his real amp, make him choose the tonestack type and pot tapers and then run the trainer that will tweak all internal parameters to make the model sound as close as possible to the user's amp.
And probably Cliff already has some plans regarding something like that, considering what his latest and recently granted patent is about.
 
Hey DLC86 !

Welcome to TGF :)

I agree. NAM is amazing. I have messed with it too and concur with your results / conclusions :)

Some "side" comments ......

Of the currently available products to use, Tonex is the clear leader - see here for a Real Amp vs QC Capture vs Tonex Capture each Null tested against the real amp and each other:- https://youtu.be/Avhih6805YU?t=1157 - the KPA is a long way off on the null test compared to the real amp .... the QC is a bit better ..... but the Tonex is in a league of its own compared to the real amp and the KPA and QC.

When I bought my KPA, I clearly remember many posts from Christoph on the KPA private forum about people "complaining" that their real amp profiles were excellent but when they tried to profile another modeler or software, the results were much worse - Christoph repeatedly confirmed this to be true and explained / stated many times that *all* his design and testing for the KPA was done with real amps and that it definitely *would not* properly capture other modelers or software emulations ..... I don't know for sure, but I strongly suspect this is exactly the same situation with the Tonex environment (?)

Guitarjon [ SDS ] recently posted a Y/T video where he Tonex Captured his JCM 800 ... and the results were quite literally indistinguishable ... he stated that even he could not hear or tell the difference

Is NAM "the best" via null tests now ..... yep !

Is Tonex noticeably better than the KPA or QC .... yep !

Even allowing for all the above, is Tonex so close / identical enough that it doesn't matter .... yep !

I.m.h.o ...... the real-world "difficulty" for NAM is that without a "major backer" and "implementer" taking them onboard, the NAM "process" probably ain't going to get any traction anywhere ..... even if a "major backer" and "implementer" took it on now, we are looking at at least 2+ <-> 3+ minimum years before we *could maybe* see *some sort of hardware* that was live-useable ..... and in the current environment, and the niche MI marker, of which digital modeling is an even much smaller niche .... I just cant see NAM going anywhere in terms of functional useability for any level for musicians other than strictly "home-based-nerds" <= no offence meant to anyone.

And even then, Tonex wont stand still ... it will keep getting better .... they have already improved the algorithm for High Gain Capturing in the first week of its release ....... the QC *may* lift their overall game ... and who on earth has any idea what Christoph is going to do ? :)

So i.m.h.o .... if anyone wants the best possible Capture / Profiling Tech to use live, in the studio or on tours etc.... ... its the Tonex ..... and I think, notwithstanding the above, this is going to be the case for many years to come.

But yes ..... NAM is amazing ! :)

All the best,
Ben

PS: I.m.h.o ..... the only logical practical place for NAM is to be somehow licensed by Line 6 for inclusion into the Helix or next top tier modeler [not Helix 2 ] that they are doubtlessly working.
Still haven't tried Tonex and NAM with a real amp (hope to do it soon, sadly I sold my last amp just 1 month ago), but from my pretty ignorant on the matter pov, seems kinda strange that it would achieve more accurate results on it than on a digital sim.
For example on the NAM facebook group pretty much all NAM users achieve lower accuracy (higher ESR - error/signal ratio) with real amps due to interferences, hum and thermal noise alone, while on sims noise is pretty much non-existent except for very low quantization noise and aliasing.

You're probably right that NAM, as it currently is, doesn't have chances for being widespread among musicians, but for me it's the first choice right now, cuz my only use for captures/profiles is to have my tones in plugin format for easy and quick reamping.
If in the future some other more user-friendly plugin reaches/surpass it or if it gets licensed to some company I might upgrade, but atm I see no reason to use any other thing for my purpose, despite having already bought a tonex license.

PS: I should also add that we don't really know if a hardware version of NAM is already in the works, even though it's become somewhat famous only in these last months/weeks, I just discovered that its first version is 4 years old.
 
Last edited:
There was supposed to be a link in that post but it wasn't included for some reason, maybe I still don't have enough posts here.
I'll try again: https://patents.google.com/patent/US20200279546A1/en?inventor=clifford+chase

EDIT: nope, the link was there but not really visible, probably something to fix on the forum theme.
Yeah links not being visible is definitely a problem.

Mods, this bit of CSS would fix them to color of the header bar:

CSS:
 a.link--external {
     text-decoration: underline;
     color: #185886;
}

Anyway, the patent seems to be just the tonematching system already in Fractal. In regard to neural networks I think Cliff has said that he has gotten more accurate results with his methods than using neural networks, though it can be argued that this is coming from a biased (no pun intended!) source.
 

It's a pre and post EQ matching sandwich.
Anyway, the patent seems to be just the tonematching system already in Fractal. In regard to neural networks I think Cliff has said that he has gotten more accurate results with his methods than using neural networks, though it can be argued that this is coming from a biased (no pun intended!) source.
It looks that's maybe a bit more than that:

"
  • Applying an ESS with a long duration in comparison to the system's impulse response to a nonlinear system results in a sequence of impulse responses. These impulse response “kernels” are related to the Volterra series expansion and represent impulse responses for the power series of the input comprising y(t)=h1(t)*x(t)+h2(t)*x2(t)+h3(t)*x3(t)+. . . .
  • [0043]
    The sequence of impulse responses is such that the linear term occurs at a time delay of T (maximum correlation time) and the nonlinear terms prior to that. If the measured response y(t) is convolved with the matched filter f(t) and the impulse response of the system is short in comparison to the stimulus duration, the result is y(t)*f(t)=h(t−T) where T is the duration of the stimulus and h(t) is the impulse response of the linear system after the nonlinearity.
  • [0044]
    Prior to T the terms h2(t), h3(t), etc. can be found which can be discarded or used to further refine the amp model by correcting the shape of the model's nonlinearity. The position of each impulse response in the time series is a function of the sweep length and the ratio of the stop and start frequencies, k. The time between each kernel is proportional to the sweep length and inversely proportional to k. The various impulse responses can be extracted by simply time windowing the result."

In another paragraph it also talks about storing user amp models and a capture method guided by a "wizard" where the user can choose a base amp model, so definitely not just tonematch.

PS: maybe this part is already included in his "SpectrumTrack"
 
It looks that's maybe a bit more than that:

"
  • Applying an ESS with a long duration in comparison to the system's impulse response to a nonlinear system results in a sequence of impulse responses. These impulse response “kernels” are related to the Volterra series expansion and represent impulse responses for the power series of the input comprising y(t)=h1(t)*x(t)+h2(t)*x2(t)+h3(t)*x3(t)+. . . .
  • [0043]
    The sequence of impulse responses is such that the linear term occurs at a time delay of T (maximum correlation time) and the nonlinear terms prior to that. If the measured response y(t) is convolved with the matched filter f(t) and the impulse response of the system is short in comparison to the stimulus duration, the result is y(t)*f(t)=h(t−T) where T is the duration of the stimulus and h(t) is the impulse response of the linear system after the nonlinearity.
  • [0044]
    Prior to T the terms h2(t), h3(t), etc. can be found which can be discarded or used to further refine the amp model by correcting the shape of the model's nonlinearity. The position of each impulse response in the time series is a function of the sweep length and the ratio of the stop and start frequencies, k. The time between each kernel is proportional to the sweep length and inversely proportional to k. The various impulse responses can be extracted by simply time windowing the result."

In another paragraph it also talks about storing user amp models and a capture method guided by a "wizard" where the user can choose a base amp model, so definitely not just tonematch.

PS: maybe this part is already included in his "SpectrumTrack"
Sounds like it's using a nonlinear filter as well. Volterra filters were pretty cool in the 90s.
 
I listened to all 3 twice.

Shooting from the hip, my immediate impression was that the FM9 has more pristine high end detail that wasn't quite there in the other 2, so I preferred the FM9.

NAM is really impressive though! And these are all totally usable tones, of course.

Thanks for doing this @DLC86 🙇‍♂️

Potential bias disclaimer: I am an FM9 owner.
 
Back
Top