When do you predict the next Fractal Audio Axe iteration will drop, now that they've stated that it will be NAM-compatible? (Poll added)

[nightlight]

Adding a poll, but some of the posts really made me wonder what direction Fractal Audio might take.
I mean, look at the FM9, it's an awesome form factor on the floor like that. Dare we assume that it might be some kind of floorboard unit?

I know that the FM9 less powerful than an Axe FXIII, but if you look at the Quad Cortex, it is definitely possible to have a really tiny package with tonnes of processing power.

 

[nightlight]

Another thing I'd be really interested in is whether the NAM-compatible profiles can be edited with deep controls like the conventional amps that have been captured, and as is the case with the modeller-approach Axe FX suite of products.

 

[Frodebro]

Another thing I'd be really interested in is whether the NAM-compatible profiles can be edited with deep controls like the conventional amps that have been captured, and as is the case with the modeller-approach Axe FX suite of products.

It would be possible to do some generic, basic stuff but it wouldn’t give you results accurate to how the original amp would respond to those tweaks-unless every component to be tweaked was programmed into the algorithm for each specific amp. Which is one of the biggest differences between captures/profiles and component-based modeling. Captures are a snapshot of a specific amp at specific settings.

 

[nightlight]

It would be possible to do some generic, basic stuff but it wouldn’t give you results accurate to how the original amp would respond to those tweaks-unless every component to be tweaked was programmed into the algorithm for each specific amp. Which is one of the biggest differences between captures/profiles and component-based modeling. Captures are a snapshot of a specific amp at specific settings.

I'd never rule out something unless I knew how guitar technology gurus like Cliff or Kemper throw their hands up and say, "There's no way to do that."

It would definitely be of use in a scenario where say you bought a NAM capture from a third party and want to fine tune it to your setup, eg. the guitar, the cab, the room, etc.

I'm assuming that there would some parameter tweaking possible, as with the controversial new controls on the Kemper V2.

 

[Blix]

I like that it will have NAM, then I get to enjoy some great captures of amps not in the Axe. Like @2dor killer Friedman Plex pack.

 

[ChordInversion]

I'd never rule out something unless I knew how guitar technology gurus like Cliff or Kemper throw their hands up and say, "There's no way to do that."

It would definitely be of use in a scenario where say you bought a NAM capture from a third party and want to fine tune it to your setup, eg. the guitar, the cab, the room, etc.

I'm assuming that there would some parameter tweaking possible, as with the controversial new controls on the Kemper V2.

And this is the $64K question. I have stated my apprehensiveness about modeler makers getting into NAM for exactly this reason - if they're going to accomplish the kind oof thing you're talking about, it's going to mean a pretty significant shift of their focus. I'd prefer not to have multiple vendors all working on enhancements for NAM, and instead would like them to keep doing their modeling work.

But, as I noted before, I"m quite happy with what I have, so I guess it doesn't need to matter to me one way or the other.

 

[nightlight]

And this is the $64K question. I have stated my apprehensiveness about modeler makers getting into NAM for exactly this reason - if they're going to accomplish the kind oof thing you're talking about, it's going to mean a pretty significant ship of their focus. I'd prefer not to have multiple vendors all working on enhancements for NAM, and instead keep doing their modeling work.

But, as I noted before, I"m quite happy with what I have, so I guess it doesn't need to matter to me one way or the other.

I'm not a rocket scientist, but the ability to tweak NAM profiles without making them sound "digital" (a common problem with tweaking parameters on the Kemper) would increase the value of such a unit to the user base significantly,

 

[laxu]

Another thing I'd be really interested in is whether the NAM-compatible profiles can be edited with deep controls like the conventional amps that have been captured, and as is the case with the modeller-approach Axe FX suite of products.

No.

The big problem with parametric captures is that you would need to capture a wide variety of knob positions. Even with a typical 5-6 knob amp this ends up being massive amounts of data, which takes a very long time to capture. Add more knobs and for 10 positions on each knob you have "10^number of knobs" possible combinations of knob positions.

Neural DSP has tried to solve this by using a "good enough" randomized sample set. You can read a paper about it here: https://arxiv.org/html/2403.08559v1
If you have followed their firmware development, ever since the announcement of the TINA robot used to turn knobs and collect the information, they have put out like maybe 3-4 new amp models. So that means their approach is not very successful overall.

If Fractal were to do something like that, it would no longer be NAM. It would be something different, with its own proprietary format.

 

[metropolis_4]

Maybe I’m just in a cynical mood this morning, but all of the talk about “everything has to have NAM and I have to be able to control it just like the real amp” sounds so typical of the lack of respect our society has for software.

We feel entitled to demand software give us everything and we expect it to be free. Decades of exposure to freeware and shareware has made us expect software to be cheap or free and we have no respect for all of the hard work, innovation, and ingenuity that goes into its creation.

We look at software as a way to get things we normally couldn’t afford for free. Amp modelers are expected to include hundreds of amps at launch, and to provide more through free updates every month. We start to use expectations of how much they’ll give us for free as a metric to decide which one to buy. And we don’t expect the cost of all that software to reflect in the price of the hardware.

The whole original idea of Profiling was that you could Profile your own gigging amp so you could have a reliable facsimile of it to bring to gigs without risking the real amp. We’ve now turned that concept into a way to expect to get a bunch of expensive amps for free with software we expect to be included for free.

 

[Orvillain]

Neural DSP has tried to solve this by using a "good enough" randomized sample set. You can read a paper about it here: https://arxiv.org/html/2403.08559v1

There's actually reasoning and logic behind the randomized approach. By doing randomized samples, you get broader unbiased coverage with fewer sample sets, you minimize overfitting issues that neural networks are prone to,

The paper even says:

The dataset totals around 4.5 hours of paired audio, randomly split into 15000 training and 1000 validation examples.

Which is actually a really good amount of data for a NN to work with.

Their network approach is an LSTM recurrent network, which is also what NAM uses iirc.

So the core idea is:
Instead of sampling knob settings on a fixed grid, sample randomly across the parameter space and train a conditioned neural network. So the model learns to interpolate across the control space. There is no guarantee of physical accuracy with this process, but the behaviour will be very close.

 

[metropolis_4]

There's actually reasoning and logic behind the randomized approach. By doing randomized samples, you get broader unbiased coverage with fewer sample sets, you minimize overfitting issues that neural networks are prone to,

The paper even says:


Which is actually a really good amount of data for a NN to work with.

Their network approach is an LSTM recurrent network, which is also what NAM uses iirc.

So the core idea is:
Instead of sampling knob settings on a fixed grid, sample randomly across the parameter space and train a conditioned neural network. So the model learns to interpolate across the control space. There is no guarantee of physical accuracy with this process, but the behaviour will be very close.

That’s very interesting, the approach makes sense. I don’t know much about machine learning or neural networks, but based on my experience with AI tools my hunch would be that the ability to interpolate and the quality of the interpolation would be the current weak point in the system.

I know AI tools have a strong tendency to make bad guesses when trying to fill in the blanks between data points. They struggle to get the context right.

I’d assume that’s an area where we’ll see improvement over time.

 

[Chocol8]

I would expect with a tool like TINA and smart software you could do far better than fixed knob increments or random settings. The software should be able to determine where it can interpolate and where it needs more data, and then it can capture what it needs.

 

[Orvillain]

As I understand it, and as the paper implies, neural network interpolation is more about how you present data, rather than picking a specific type of interpolation. If you present a grid like linearized structure, then it interpolates badly. But by presenting the randomized data set, with the correct tag structure, it can more accurately perform its internal interpolation.

Very different to something like sample interpolation, where you have various approaches (linear, cubic, hermite, etc)

 

[MirrorProfiles]

Is there actually a documented use of where TINA has been used? IIRC that paper about it is quite old and I wouldn’t be surprised if it’s been used far longer than that youtube video has been active for

 

[kvik]

I wish Fractal would release a modelling API and let us build (and sell) our own amp models and effects etc. to run on top of their platform. I surmise it would create a sprawling third party ecosystem that parallels what's going on in the captures space. Extend the API into the user interface and let people create their own workflows and we're talking ultimate potential sound processing platform with nobody coming even close.

Unfortunately this is probably very far from being a reality for multiple reasons. But one is allowed to dream.

 

[Mongillo19]

Unfortunately this is probably very far from being a reality for multiple reasons. But one is allowed to dream.

That you are!

However, if some people can't make good sounding profiles of their amps (which especially DI profiles is easy). Can you imagine the garbage amps people would make

 

[Orvillain]

Is there actually a documented use of where TINA has been used? IIRC that paper about it is quite old and I wouldn’t be surprised if it’s been used far longer than that youtube video has been active for

Oh almost certainly. I wouldn't be surprised if it was used for most of their plugins to be honest. At least the 'X' ones.

 

[laxu]

I wish Fractal would release a modelling API and let us build (and sell) our own amp models and effects etc. to run on top of their platform. I surmise it would create a sprawling third party ecosystem that parallels what's going on in the captures space. Extend the API into the user interface and let people create their own workflows and we're talking ultimate potential sound processing platform with nobody coming even close.

Nah, that never works. The number of people developing effects would be minimal.

Fractal did actually allow people to make their own peripherals and user interfaces for their past generations by releasing a MIDI Sysex specification. The current gen system is not that far off from that, but enough that you need to reverse engineer it if you want to make something.

 

[nightlight]

Maybe I’m just in a cynical mood this morning, but all of the talk about “everything has to have NAM and I have to be able to control it just like the real amp” sounds so typical of the lack of respect our society has for software.

We feel entitled to demand software give us everything and we expect it to be free. Decades of exposure to freeware and shareware has made us expect software to be cheap or free and we have no respect for all of the hard work, innovation, and ingenuity that goes into its creation.

We look at software as a way to get things we normally couldn’t afford for free. Amp modelers are expected to include hundreds of amps at launch, and to provide more through free updates every month. We start to use expectations of how much they’ll give us for free as a metric to decide which one to buy. And we don’t expect the cost of all that software to reflect in the price of the hardware.

The whole original idea of Profiling was that you could Profile your own gigging amp so you could have a reliable facsimile of it to bring to gigs without risking the real amp. We’ve now turned that concept into a way to expect to get a bunch of expensive amps for free with software we expect to be included for free.

What are you saying? NAM is free anyway. No one is asking NAM to improve its feature set, and the Fractal product is likely to cost thousands of dollars.

I was talking about Fractal Audio giving users the option to tweak captures. Doesn't have to be an unlimited set of controls, I'd really appreciate something like the "Definition" control on the Profiler, which - let's face it - every profiler maxed out for heavy tones because it made the profiles sound tighter/better.

 

[Mongillo19]

Oh almost certainly. I wouldn't be surprised if it was used for most of their plugins to be honest. At least the 'X' ones.

This was just a few weeks ago. Sure seems like the Nameless plug uses/used TINA. Probably when they made it PCOM and gave it the X update.