Quad Cortex plugin support officially goes from "soon" to "eventually"

[JiveTurkey]

So… they still can‘t manage to port their plugins to the hardware, but are planning to make a plugin version of the hardware‘s software?

 

[GTR37]

Friend of mine brags about marrying a Miss P. Enis

No joke

Hopefully he has found true apenis

 

[signalpath]

The aspirated h went awol.

 

[fretworn]

Hopefully he has found true apenis

He was actually a business partner of mine… a truly talented sw engineer and business guy.

Brilliant fella who pioneered how to attach relational databases to the Internet (like 1990; the public at large starts to see such work 12+ years later at best.)

Crudest mutha’fukka you’ll ever meet. A former carrier sailor from LA (Lower Alabama). He read the book “Everybody Poops” to my kids when they were very young.

His farts are EPIC… whole ‘nother thread. Airplane story ‘fer reelz.

If you threatened to throw down on him, he would ask you to pull out your dick, and would exclaim loudly that “he would show you something you’ve never seen before… The other end.” Lol.

Solid guy.

(He also showed me what alt.binaries were early on. Some things you never unsee… For a lifetime.)

 

[DLC86]

FWIW:
If you were to do interpolation-based black-box it's an N-dimensional interpolation table, where N is the number of controls, and each entry in the table is a vector. For example, if there are only two knobs then it's a two-dimensional lookup table and you interpolate linearly between adjacent vectors.

For interpolation to work you typically need more than a few samples. I would consider 5 samples per control the bare minimum. For a control that goes 0-10 that means samples at 0, 2.5, 5, 7.5 and 10. I'm not sure how well that coarse of a sampling would work though. It depends on the locus of the coefficients vs knob position.

So for an amp with, say, Gain, BMT, Presence, Depth and Master you have seven controls. That means your table is 7-dimensions which, for five samples is 5^7 = 78,125 entries.

If it takes a minute to learn each combination of settings that's 54 days. And that's assuming you have a robot that can adjust these controls for you accurately, nothing ever malfunctions and the power doesn't go out. I've yet to see a robot that can be programmed to turn amp knobs but I'm sure it could be done. It would likely be very expensive.

If an amp has switches then that's another dimension but that dimension only has two (or three if it's a three-way switch) elements. So if our hypothetical amp has a three-way switch it would now take 162 days to learn it. If it has two three-way switches then it's 324 days. Etc.

Now we get into data storage. Let's say there's 1K coefficients that are required for each vector. This means you need to store 78M coefficients (for each switch position).

Given the above the interested reader can probably draw conclusions about the techniques being used for adjustable amp models.

This is why I think a hybrid approach could be the best, machine learning applied to underlying amp models which already have all/most amp topologies and tonestack types and the training only tweaks existing parameters on the model, after the user chooses the same/closest model and sets the knobs' positions and tapers to be the same as on the real amp... if that's even feasible.

Cuz there's one (well, actually more than one) thing that I still haven't grasped though regarding this: we know that white box modeling requires taking measurements along the whole circuit of the amp to fully replicate the behaviour of every stage and component. But these single component/stage behaviours somehow have an effect on the output (otherwise it wouldn't make sense to model them), so in theory, even if it wouldn't be an easy task for sure, it would be possible to extrapolate e.g. the power tube bias (or any other advanced/internal parameter) value just by comparing the input and output, with an appropriate test signal and long enough training time. Or is it impossible for some reason?

I've yet to see a robot that can be programmed to turn amp knobs but I'm sure it could be done. It would likely be very expensive.

Or install motorized pots/faders in the amp with the same resistance and taper of the original ones. But that wouldn't be exactly "user friendly"

 

[imperator]

This is why I think a hybrid approach could be the best, machine learning applied to underlying amp models which already have all/most amp topologies and tonestack types and the training only tweaks existing parameters on the model, after the user chooses the same/closest model and sets the knobs' positions and tapers to be the same as on the real amp... if that's even feasible.

I would model the the preamps with NNs, then just use traditional modeling elsewhere. This assumes you can isolate a large part of the preamp from the controls.

 

[DLC86]

I would model the the preamps with NNs, then just use traditional modeling elsewhere. This assumes you can isolate a large part of the preamp from the controls.

And that's the problem, usually the tonestack is a part of the preamp and you can't really isolate it without taking the amp apart. The other way around could be more feasible

 

[imperator]

And that's the problem, usually the tonestack is a part of the preamp and you can't really isolate it without taking the amp apart. The other way around could be more feasible

I always thought t the tone stack usually sat between the preamp and Poweramp or before it like on a mark?

 

[laxu]

I always thought t the tone stack usually sat between the preamp and Poweramp or before it like on a mark?

It varies on an amp to amp basis. It's early in the circuit on Fender type amps and later in Marshall type amps. You can't easily separate it out because it has a big effect on how much gain and with what kind of frequency response goes to the later preamp stages as well as the poweramp.

Mesa Mark series have one tone stack early on (the knobs) and one later on (graphic EQ). The effect of those two sets of controls are quite different.

 

[DLC86]

What @laxu said!

 

[imperator]

It varies on an amp to amp basis. It's early in the circuit on Fender type amps and later in Marshall type amps. You can't easily separate it out because it has a big effect on how much gain and with what kind of frequency response goes to the later preamp stages as well as the poweramp.

Mesa Mark series have one tone stack early on (the knobs) and one later on (graphic EQ). The effect of those two sets of controls are quite different.

Certainly looks feasible on a lot of amps https://robrobinette.com/How_the_Marshall_JCM800_Works.htm

That being said, I think at this point fractal has so much stuff that's done, their approach of gradient boosting seems right.

 

[Lysander]

Shifting the discussion back on track ...

Plugin Development Update - March 2023

Thanks for the update on your direction. Your goals are important, both Apple Native (M1/2) and the QuadCortex. I respect your openness and appreciate your direction. Being a Plugin owner running a Mac M1, I know how important this is. Running the plugin in Rosette makes it a bit of a hog...

unity.neuraldsp.com

 

[norminal]

View attachment 4847

Apple Silicon shouldn’t even be part of the discussion. Apple gave software developers more than a generous notice to get their code written, offered development units, etc. It’s been what, two years now? Annnd we’re already on M2

Quit passing the blame to Apple, Neural DSP.

 

[Lysander]

Apple Silicon shouldn’t even be part of the discussion. Apple gave software developers more than a generous notice to get their code written, offered development units, etc. It’s been what, two years now? Annnd we’re already on M2

Quit passing the blame to Apple, Neural DSP.

Also, i feel like i'm taking crazy pills, because this...

"To clarify this point in the original post, we will be allocating a large portion of the remaining resources that would be focusing on new products to prioritise Native Apple Silicon support, but this will not affect the rate at which we continue to work on Quad Cortex or Plugin Compatibility."

...is the exact opposite of what Francisco Cresp stated a week ago.

 

[DrewJD82]

Hahahaha they’re damned if they do, damned if they don’t-

They try to communicate after customers complained about the lack of communication but can’t line up their own communication.

“Hey guys, let me clarify some things-

,

Got it?”

Also, I’ve noticed Doug isn’t the one doing these updates. I wonder if someone sat him down and said “Please let us handle communication as our ‘No problem’ and ‘Absolutely!’ check book is deplete and the bank’s ain’t giving us anymore.“

 

[mbenigni]

Shifting the discussion back on track ...

Plugin Development Update - March 2023

Thanks for the update on your direction. Your goals are important, both Apple Native (M1/2) and the QuadCortex. I respect your openness and appreciate your direction. Being a Plugin owner running a Mac M1, I know how important this is. Running the plugin in Rosette makes it a bit of a hog...

unity.neuraldsp.com

View attachment 4847

I'm reading this out of context, but the quoted text and subsequent paragraph seem very much at odds. "development and QA teams will halt work..." vs. "will not affect the rate at which we continue to work..."

What's Cortex Control? Is that what they're calling their editor application now?

 

[DrewJD82]

I'm reading this out of context, but the quoted text and subsequent paragraph seem very much at odds. "development and QA teams will halt work..." vs. "will not affect the rate at which we continue to work..."

What's Cortex Control? Is that what they're calling their editor application now?

Come on, man! They’ve explained it twice in one week now! What’s not to understand?

 

[James Freeman]

They have such a charming way with words...

 

[mbenigni]

Come on, man! They’ve explained it twice in one week now! What’s not to understand?

They explained it twice in one sentence. Three times. Never the same way twice.

 

[mbenigni]

Return it. At this point; you've done your due diligence.

Just realized the downside of ordering direct from IK versus e.g. Sweetwater. Can't find anything on return policy beyond the "All sales are final" text in my email receipt. Not surprising considering so much of their revenue is per software attach rate. I just wrote customer service, but I've been sitting on this thing for about three weeks now, and I'm probably pushing my luck.

EDIT: Hold the proverbial phone. Jason at customer support got back to me about 30 seconds after I logged the ticket, so maybe this will turn out OK...