QC, ToneX and the future of Fractal

[Dave Merrill]

Except that on many amps all those knobs interact with each other.

I still don't think that approach will solve it. Although it would probably improve things.

Yeah.

To do this right you'd need to "record" every combination of settings for every control.

That's a lot, especially if you model the "controls", ie block parameters, in as much detail as Fractal does.

 

[ruso]

Except that on many amps all those knobs interact with each other.

I still don't think that approach will solve it. Although it would probably improve things.

My guess is the workaround would be capturing a smaller number of knob positions and then using, dare I say, machine learning, to “fill in the gaps.”

 

[lqdsnddist]

Except that on many amps all those knobs interact with each other.

I still don't think that approach will solve it. Although it would probably improve things.

I guess if there was time one could turn one knob while it ran measures and then you set it to 0, or 5 etc, turn the other knobs etc, after a few dozens measures perhaps you'd have something that could cover most of the range and somehow interpolate some in between variables ?

Then again, if one could have say a dozen different dialed in amp tones for a given amp at the click of a mouse, is there a big need to be able to tweak significantly from the captured tone ? You could go and adjust EQ, pre and master et al, or you could just click "edge of breakup" profile and get back to jamming

 

[peteri]

My guess is the workaround would be capturing a smaller number of knob positions and then using, dare I say, machine learning, to “fill in the gaps.”

No disrespect to this post, but as an example - that phrase 'machine learning' is really a lot of marketing speak in 99% of times it's cited.

And I believe that Tonex is a case in point, machine learning cannot miracle up interactions from nothing - to quote Wikipedia 'Machine learning algorithms build a model based on sample data, known as training data, in order to make predictions or decisions without being explicitly programmed to do so.'

There's no way it can be applied to 'guessing' how two pots with random tolerances and behaviours can interact with a large complex circuit full of the same wandering tolerances (and remember those tolerances will interact differently day on day given temperature, humidity and voltage variance.

No disrespect intended

IF Tonex does apply machine learning (and that's a big if), I believe effectively what they're describing is basically how it interpolates between the various sample points taken when you profile the amp, to profile the non linear behaviour of the machine (i.e. the amp) as a whole. It's a whole other can of worms to cover the gap between knob positions. From my playing with Tonex (which I enjoy btw), I think it's likely doing that with the changes of behaviour due to input level (i.e. response to volume) which is way better than others I've tried - but not much more than that

 

[ruso]

No disrespect to this post, but as an example - that phrase 'machine learning' is really a lot of marketing speak in 99% of times it's cited.

No disrespect to your response, but I intentionally used quotation marks to get across the ridiculousness of the whole “AI/Machine Learning” marketing that is flooding the MI industry.

 

[northern_fox]

Yeah.

To do this right you'd need to "record" every combination of settings for every control.

That's a lot, especially if you model the "controls", ie block parameters, in as much detail as Fractal does.

That's how NDSP makes the plugins and models, but via automation

And as I understand it, it's a little more complicated than simply blending a few captures after the fact. It's likely a long process that involves a single model instead of multiple small ones.

 

[unix-guy]

That's how NDSP makes the plugins and models, but via automation

And as I understand it, it's a little more complicated than simply blending a few captures after the fact. It's likely a long process that involves a single model instead of multiple small ones.

Do they use some sort of robot to control all the knobs?

Imagine amps that have 10-12 knobs on them. The near infinitude of possible knob setting combinations would take a LOT of effort!

 

[nicolasrivera]

Before anything, from my experience dealing with many new comers to the Fractal world locally, they do struggle with the learning curve a lot. Line 6 and QC win them over with, more then anything else, a way simpler user interface and intuitive features.

So, before talking about captures, touch screens etc, Fractal may very well overhaul the UI with something that don't frustrate the new users and invites new ones, many musicians just don't have the time to deal with many things because of a tight schedule.

my 2cents.

 

[northern_fox]

Do they use some sort of robot to control all the knobs?

Imagine amps that have 10-12 knobs on them. The near infinitude of possible knob setting combinations would take a LOT of effort!

Yes actually, it's mechanically automated. That said, I'm quite sure it's not every combination of every knob and switch, which would be simply too much time and too big a model.

 

[Greg Ferguson]

Except that on many amps all those knobs interact with each other.

I still don't think that approach will solve it. Although it would probably improve things.

Agreed. Every combination of the knobs would have to be captured or it’d be the same old situation.

 

[Greg Ferguson]

My guess is the workaround would be capturing a smaller number of knob positions and then using, dare I say, machine learning, to “fill in the gaps.”

ChatGPT for profiling. Great.

 

[GlennO]

Do they use some sort of robot to control all the knobs?

Imagine amps that have 10-12 knobs on them. The near infinitude of possible knob setting combinations would take a LOT of effort!

I take you've never sat through a sampling session. Much of the technology you rely upon is based on a someone making a lot of effort .

 

[SacredGroove]

Yeah.

To do this right you'd need to "record" every combination of settings for every control.

That's a lot, especially if you model the "controls", ie block parameters, in as much detail as Fractal does.

I imagine it could be done similarly to how drum software can detect a specific drum in a kit while the whole kit is being played, as well as accompanied by other instruments.
If the software already models how all the knobs interact with each other, based on the specific section of the amp (i.e. input gain, tone stack) then there shouldn't be any need to perform every possible combination of knob settings. Just a guess though.

 

[unix-guy]

I take you've never sat through a sampling session. Much of the technology you rely upon is based on a someone making a lot of effort .

I have not... Doesn't sound like a good time

 

[northern_fox]

I imagine it could be done similarly to how drum software can detect a specific drum in a kit while the whole kit is being played, as well as accompanied by other instruments.
If the software already models how all the knobs interact with each other, based on the specific section of the amp (i.e. input gain, tone stack) then there shouldn't be any need to perform every possible combination of knob settings. Just a guess though.

The trick here is that it's not really 'modeling' anything. What's both good and not good about ML stuff l like this is that you're worrying less about what's actually happening, and more just on measured input vs output. You can strategize about what to feed it for training and the training parameters, but ultimately don't really have to worry about being specific about which components are interacting with which as long as you give it good data. It will figure out the rest, in theory. You don't need to really know anything about the gear or how it's actually working to get a good result - but I think it would certainly help in determining how to spend time.

 

[yeky83]

Yes actually, it's mechanically automated. That said, I'm quite sure it's not every combination of every knob and switch, which would be simply too much time and too big a model.

Source? Where have they said they have automated bots turning knobs and they’re making captures along the way?

 

[northern_fox]

Source? Where have they said they have automated bots turning knobs and they’re making captures along the way?

I don't want to link here really but Doug talks about it in one his interviews and it's been discussed before. It's not captures along the way though, it's one big model. It's a specific kind of ML process for this (I forgot the name, not my background) but it's a way of introducing more variables into the model to learn. NDSP didn't invent this, but they are applying it to get parametric ML models. Actually, NAM is capable of something similar (or at least, Steve had done it once).

 

[ruso]

 

[northern_fox]

I mean, yeah - none of us knows for sure what they're doing, obviously. To be fair, that post doesn't actually indicate one way or another - just acknowledges the difficulty in covering the broad space. But one way or another, parametric ML capturing is already technically possible to some extent. I'd be surprised if we don't see more of it, or hybrid component / ML solutions in the near future (not necessarily from NDSP).

 

[unix-guy]

View attachment 117066

Yes, that's exactly what I was thinking when I posted above.

And even that, in my opinion, is still relying heavily on interpolation, which may or may not be accurate.

Between 1 and 10 is not 10 positions, but a large range depending on how fine you control the knob... And probably further complicated by things like Pot values, linear vs log10 vs log20, etc...