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

IIRC he's said their plugins don't have a singular approach to modeling and they change the methods as they develop them, adding in black boxes as they see fit.

He keeps saying this kind of stuff in interviews:
"We record signals for a few hours, and then we train the model on those... On our own servers, it would be four to six hours of recording, and then eight hours of training."

So they can pop out an amp model in 14 hours and it's all automated... Really? Where's all the amp models?

And it's contradictory to the issue of knob permutations that DC himself laid out:
Let's say 7 knobs on an amp, 10 positions per knob. That's 10^7 permutations.
A capture signal on the QC takes ~1 minute. So play a 1 minute signal 10^7 times, that's 7000 days.
How are they getting it done in supposedly 4-6 hours of recording?
I would guess that they need less overall signals for different permutations of knob positions. Or something simplified like min, mid, max or every 2-3 steps or something. Or they capture the effect of the knob being sweeped from 0-10.

In any case they have managed to model pretty complex amps like the JP2C in the QC so whatever they do it clearly does work.
 
The sausage quote was in regards to a comment saying IK ToneX will cut into QC and Kemper sales. Person goes on to say FAS will be their digital gear of choice if they do go that route.
 

Attachments

  • Screenshot_20230306-135210~2.png
    Screenshot_20230306-135210~2.png
    20.5 KB · Views: 54
  • Screenshot_20230306-135142~2.png
    Screenshot_20230306-135142~2.png
    36.5 KB · Views: 52
I would guess that they need less overall signals for different permutations of knob positions. Or something simplified like min, mid, max or every 2-3 steps or something. Or they capture the effect of the knob being sweeped from 0-10.
I dunno anything about ML, but those suggestions don't make sense to me. Less test signal length, how much less? Less permutations, how do they train accurately with no training data? Sweep knobs, huh?
In any case they have managed to model pretty complex amps like the JP2C in the QC so whatever they do it clearly does work.
The issue isn't whether they can model amps or not. It's whether they actually have an automated process of pumping out an amp model a day as claimed. And if they do, why don't they?
 
I had an operation called SMILE. They laser a small hole into your eye and through there cut off a piece of the lens, then use tools to extract it. This is how the extraction part looks, like literally reminded me of windshield wipers, but on my eyes.

arnold schwarzenegger windshield wipers GIF
:rofl
I had lasik done in October 2020.. Best 5 grand I ever spent!
 
On Leon's discord a former employee of neural was chatting about gear. They mentioned how if he goes digital is will certainly not be NDSP. Now sure there could have been some divide if said person was let go etc but they mentioned how since they "know how the sausage is made" at NDSP, he will be avoiding their products.

Could be something, could be nothing
Yeah impossible to tell with nothing more to go on.

I can understand it though. I worked as a software consultant for <large financial sector company> and their software running on <redacted> was terrible, to the point that I will never use their services. The development was spearheaded by a few capable developers and a lot of not very capable ones hired by a large software consulting agency.

Then there's also seeing your own work and you go "Ahhh, I could have done that so much better. Why did I do it like that?" when you have more experience under your belt. I cringe at some of the code I wrote back in the day. In 10 years I will probably cringe at code I wrote today.
 
i have no doubts there's at least plans to get that done. But all too apparently, it's not just as easy. Or their resources are too limited. Either way, it's frustrating for their users - and having to wait longer and longer certainly doesn't help.
Yes, like any other purchase its probably best to decide whether what it does today is worth it or not. I've only ever viewed the plugins as new models for the device, so I don't care much as long as the device continues to get new models and captures. I don't have a bunch of presets I want to port. I've never gotten how bent people are about the plugins.

That said, I think the ToneX has changed the value proposition for the QC quite a bit, though. The caps and the flexibility/power/UI combined with the form factor have been the thing with the QC, IMHO. It's really awesome in practice if you take it at face value for what it is right now and a are amenable to captures... But... with ToneX, the Idea of like a Stomp XL + ToneX suddenly looks like a good to great alternative to a QC.

I really don't think a bunch of new models from the plugs would change any of that for me either way.
 
I dunno anything about ML, but those suggestions don't make sense to me. Less test signal length, how much less? Less permutations, how do they train accurately with no training data? Sweep knobs, huh?

The issue isn't whether they can model amps or not. It's whether they actually have an automated process of pumping out an amp model a day as claimed. And if they do, why don't they?
Those are the questions aren't they. As you put in numbers, doing lengthy sweeps at every possible knob permutation would take years so there has to be shortcuts taken to achieve results that work.

Less permutations could possible allow for a "good enough" result - I mean real pots are often not e.g properly logarithmic either but an approximation.

You could also combine techniques where you have the ML stuff paired with more traditional signal shaping. I assume there's plenty of novel approaches they could try to get around having excessive training time while still having an amp model that behaves similar to the real thing.
 
Yes, like any other purchase its probably best to decide whether what it does today is worth it or not. I've only ever viewed the plugins as new models for the device, so I don't care much as long as the device continues to get new models and captures. I don't have a bunch of presets I want to port. I've never gotten how bent people are about the plugins.

That said, I think the ToneX has changed the value proposition for the QC quite a bit, though. The caps and the flexibility/power/UI combined with the form factor have been the thing with the QC, IMHO. It's really awesome in practice if you take it at face value for what it is right now and a are amenable to captures... But... with ToneX, the Idea of like a Stomp XL + ToneX suddenly looks like a good to great alternative to a QC.

I really don't think a bunch of new models from the plugs would change any of that for me either way.
Have you tried the Tone-X yet?
 
Less permutations could possible allow for a "good enough" result - I mean real pots are often not e.g properly logarithmic either but an approximation.
It's not about pot values and their taper though, is it. The system behaves differently depending on where the knobs (which are interactive and may be placed anywhere in the gain stages) are set. How's the NN supposed to learn the system's response at say BMT: 1,9,1 if it doesn't have that data?
You could also combine techniques where you have the ML stuff paired with more traditional signal shaping. I assume there's plenty of novel approaches they could try to get around having excessive training time while still having an amp model that behaves similar to the real thing.
Yeah, I'm guessing what they do in actuality is combine black box & white box techniques, which other companies do as well already.
 
Those are the questions aren't they. As you put in numbers, doing lengthy sweeps at every possible knob permutation would take years so there has to be shortcuts taken to achieve results that work.

Less permutations could possible allow for a "good enough" result - I mean real pots are often not e.g properly logarithmic either but an approximation.

You could also combine techniques where you have the ML stuff paired with more traditional signal shaping. I assume there's plenty of novel approaches they could try to get around having excessive training time while still having an amp model that behaves similar to the real thing.
Maybe it’s just interpolation between a limited number of Captures.
 
Have you tried the Tone-X yet?
I've given the S/W a half hearted go on my Syn30 and a couple modules. At first blush it seems to be in the same class as the current high end (QC, KPA) and am giving it the benefit of the doubt based on that plus current clips/comparisons/user reports. Also assuming the H/W is the same or better in terms of tones/feel/latency.

I am doing a massive reconfiguration of my home studio stuff so have not really been in any kind of position (or motivated) to do much in terms of capture A/Bs.
 
He’s said a number of things over time. Some of which possibly seem to be contradictory at times.
Definitely. He wouldn't even have to necessarily be intentionally deceptive, he may not even know a lot of the technical details with 100% certainty. He is just the CEO after all. I doubt he has personally implemented much of the FW.
 
It's not about pot values and their taper though, is it. The system behaves differently depending on where the knobs (which are interactive and may be placed anywhere in the gain stages) are set. How's the NN supposed to learn the system's response at say BMT: 1,9,1 if it doesn't have that data?
Yeah, but you can probably limit the permutations needed so it's not like say every knob at every numeric position but something like 1-3-1, 1-6-1, 1-9-1, 3-3-1, 3-6-1, 3-9-1, 6-3-1 etc. Maybe capturing this interaction could also be achieved with a shorter impulse so it takes less time to measure each combination.

Or they can skip all that and do it completely differently. Take the BluGuitar Amp 1 Mercury Edition. I feel its 3 all analog drive channels can emulate nearly any Marshall-based amp ever made just by varying the side tone knob and the 3-band EQ. The EQ is low shelf, 600 Hz midrange and high shelf filters. The side tone knob blends between preset tone stacks on the amp. It works remarkably well in my experience.
 
Maybe it’s just interpolation between a limited number of Captures.

I think that's the next logical step in the capturing realm. Based on whatever how many captures, interpolation could serve to suit any inbetween values.
But for obvious reasons, even if that worked, you'd likely need a whole number of captures to get close to the real deal. And once you had that whole number of captures, you'd possibly not need all the interpolation to work anymore (in the end, it'd only save you from dealing with multiple captures).

Whatever, I think as a first step, the entire EQ game should be upped considerably. With a mere handful (depending on the amp type perhaps even just one or two) captures, proper pre- and post-EQing could take you *very* far. I mean, anyone who ever fooled around with a fully parametric midband in front of whatever drive stage, will instantly know how much you can alter the drive characteristics.
So, give me two fully parametric EQs (or maybe just parametric mids and tuneable low/hi shelves) and allow me to assign any of those to a BMT tone stack for easier access.
And then let me save any of these tonestacks as "partial" presets, so I can add them to any capture quickly.

Another idea to get a lot more out of one capture would be, say, "macro" controls. A treble knob couldn't just raise trebles but possibly also slightly raise your input gain. Or just cause a slight mid-boost, too. That way, you could sort of mimic interactive tone stack bands.

And finally, add input level as a modifier/modulator. One thing I sometimes notice (regardless whether it's a component modeled thing or a KPA (the only capturing device I have at least a little bit of experience with)) would be that the dynamic behaviour of the models is different from the original. Sometimes they don't clean up well, sometimes they get shrill when you hammer the input, etc. - all that could sort of easily be adressed with input level as a modulator. Want more clean up when turning down your guitar volume? Well, just modulate gain via input level. The possibilities are almost endless.

Needless to say, I want all these on component based modelers, too.
 
The side tone knob blends between preset tone stacks on the amp.

That's pretty much along the lines of what I wrote. And it's something not taken care of much in the current lineup of capturing devices. The Tonex offers pre or post EQ and that was it. But what about pre-mids and post-B/T?
 
Definitely. He wouldn't even have to necessarily be intentionally deceptive, he may not even know a lot of the technical details with 100% certainty. He is just the CEO after all. I doubt he has personally implemented much of the FW.
I think it just adds to some of the general confusion. Even if Doug is not familiar with the details he should be up on meta stuff. I’ve seen the models described both ways. In that interview it reads like they made a transition at one point. Though to me the modeling method is largely academic as long as the results are there.
 
Yeah, but you can probably limit the permutations needed so it's not like say every knob at every numeric position but something like 1-3-1, 1-6-1, 1-9-1, 3-3-1, 3-6-1, 3-9-1, 6-3-1 etc. Maybe capturing this interaction could also be achieved with a shorter impulse so it takes less time to measure each combination.
OK, let's do napkin paper math with low estimates.
1, 3, 6, 9 positions for 7 knobs, 4^7. Let's say the test signal takes 10 seconds. That's still 46 hours.
 
Back
Top