How does a Quadcortex capture function work?

[Christoper]

This may be a fairly broad question, because I know black-box input/output analysis is quite a complex field.

How does a Quadcortex or similar type of DSP device capture parameters? I was thinking about this specifically with regards to interactive controls. It would be simple enough to model input gain for example, just lower the volume of the signal going into the capture, but that's not how a lot of pedals work. A Rat for example, the lower the gain control the less bass the signal has. ie, boosting a Rat that's set to low gain will sound different than having a Rat by itself that is set to have the same amount of distortion.

I have never used a Quadcortex because I am poor, but I would assume that to capture a pedal with three interactive controls you would need to at the very least let the pedal run 3^N captures where N is the number of different points along each knobs rotation you want to capture. They probably do something more automatic though

[potul]

As far as I know, they don't capture parameters. When you capture an amp or pedal, you capture a snapshot of it in its current control positions.
Then, you can apply some tweaking after the capture (input/output volume, EQ, etc...), but these are not captures of the real device controls.

[marcelomd]

+1 for capturing a snapshot of the current settings.

I've seen Kemper profiling amps on youtube. Looks like a it's capturing a set of impulse responses. Sounds like an alien calling home.

I've also seen Billy Sheehan and Line 6 talking about their colaboration. They mention modeling the circuit. Probably they break the circuit into blocks (each stage, eq between stages, etc.) and get impulse responses for said blocks (as opposed to something like SPICE modelling).

[taudio]

Most of these approaches are using some sort of non-linear system identification.

As you noted, they treat the amp to be captured as a "black box" (as in we can't see what is inside it) and probe it with a variety of different input signals (over level, frequency etc.) while capturing the output signal.

A model is derived from an analysis of the input and output signals. By probing with a wide range of input signals the nature of the overdrive / distortion and the frequency response can be modelled. My guess is that any tone controls are simply put on top of the model (since most modeling approaches don't ask you to adjust the tone controls as part of the capture process).

A good example of how this is done can be found here.

They use PyTorch (likely with a long term short term memory neural network) to train a model that emulates the captured amp (as represented by the input and output signals).

[Christoper]

That makes a lot of sense. So the best capture of my Rat would be to set the gain, tone, and volume controls wide open. Then I would design a set of user controllable input and output filters, and then my program would look like INPUT FILTERS -> RAT MODEL -> OUTPUT FILTERS. Keeping on with the Rat as an example, I would design a volume cut control on the input that also simulates how the op amp feedback loop HPF operates, I would also create a filter model on the output that emulates the Rat tone control, and finally a linear volume control on the output.

[potul]

I wonder if QUAD cortex has somthing like a dynamic capturing. I just saw this announcement, so it looks like they are working on a way to capture multiple instances of the same amp at different settings:

[ElectricDruid]

I wonder if QUAD cortex has somthing like a dynamic capturing. I just saw this announcement, so it looks like they are working on a way to capture multiple instances of the same amp at different settings

It starts to look like "sledgehammer to crack a nut" stuff though, doesn't it? There's no need to throw this kind of power at stuff that we already know and understand, like basic tone control networks in front of the non-linerities.
Still, like that "Every problem looks likes a nail to a man with a hammer", every problem looks like a nut that needs a sledgehammer to a man with a sledgehammer!! And there *is* honestly a sense in which just throwing more computing power at it is the simplest and easiest way to solve the problem, so I'm not exactly *against* it...I suppose I feel it's just a bit of a *waste* is all!!

[FiveseveN]

it looks like they are working on a way to capture multiple instances of the same amp at different settings:

Looks like CGI to me. Do you think they built real Plini and Gojira amps for a 2 minute video?! Why would they make them look like that for the purpose of capturing audio?

[cordobes]

I don't think he meant to say that the video literally shows what they are doing, because it is obviously cgi, and quite abstract.
It would be interesting to see some white paper, Strymon style, especially on IR handling, transitions, nonlinear models, reparametrization, etc, with the novelty of plugins.
I remember reading some time ago that they got patents around more efficient convolution algorithms and some other techniques.
Anyway, looking at other posts with diy developments around 32-bit architectures, achieving a good sound, nowadays, is trivial.

[marcelomd]

That makes a lot of sense. So the best capture of my Rat would be to set the gain, tone, and volume controls wide open. Then I would design a set of user controllable input and output filters, and then my program would look like INPUT FILTERS -> RAT MODEL -> OUTPUT FILTERS. Keeping on with the Rat as an example, I would design a volume cut control on the input that also simulates how the op amp feedback loop HPF operates, I would also create a filter model on the output that emulates the Rat tone control, and finally a linear volume control on the output.

Modeling the Rat with wide open tone and then adding a digital filter after it might work ok. The Rat's tone control is linear.

Modeling the Rat with wide open gain and then adding a volume control before it, not so much. Big signal into low gain sounds different than small signal in into high gain.

If you simulate the circuit you'll see the frequency response of the  gain stage changes with gain. Then you add all the non-linearities of the particular opamp being overdriven and clipping itself.

Maybe it sounds good, who knows? But it is not an accurate model.