NDSP Quad Cortex

[Mongillo19]

Yeah so did I.

Hopefully we get some answers

 

[2dor]

Hopefully we get some answers

SOON

 

[MirrorProfiles]

wonder how much Mesa’s incorrect schematics and badly laid out, poorly labelled PCB’s drove them to this

 

[laxu]

End-to-End Amp Modeling: From Data to Controllable Guitar Amplifier Models

If any of you smart people want to see what TINA is all about

Interesting stuff. I am not going to say I understand much about the equations and math involved, and it still seems pretty vague about how the control positions map into the system.

The random sampling is an interesting way to get around requiring a massive number of data points and also to reduce wear and tear on the amp controls.

The one flaw I can see with Neural's approach is that it doesn't account for different speaker loads altering the behavior of the amp. Each model would always just work as if it was connected to whatever loadbox or cab they used for generating the data set.

Honestly this is probably good enough for most people because most don't mind using a single reactive load or attenuator with their whole amp collection. But it would give component modeling an advantage by being more adaptable to what happens with e.g pedals connected, or different cabs connected.

 

[2dor]

Interesting stuff. I am not going to say I understand much about the equations and math involved, and it still seems pretty vague about how the control positions map into the system.

The random sampling is an interesting way to get around requiring a massive number of data points and also to reduce wear and tear on the amp controls.

The one flaw I can see with Neural's approach is that it doesn't account for different speaker loads altering the behavior of the amp. Each model would always just work as if it was connected to whatever loadbox or cab they used for generating the data set.

Honestly this is probably good enough for most people because most don't mind using a single reactive load or attenuator with their whole amp collection. But it would give component modeling an advantage by being more adaptable to what happens with e.g pedals connected, or different cabs connected.

That load piece can somewhat be tackled with some IRs that make up the delta of the load they modeled the amp with vs other cabs connected to the same amp.

 

[MirrorProfiles]

The one flaw I can see with Neural's approach is that it doesn't account for different speaker loads altering the behavior of the amp. Each model would always just work as if it was connected to whatever loadbox or cab they used for generating the data set.

Honestly this is probably good enough for most people because most don't mind using a single reactive load or attenuator with their whole amp collection. But it would give component modeling an advantage by being more adaptable to what happens with e.g pedals connected, or different cabs connected.

I guess they could seperate capturing the preamp and poweramp (to an extent anyway), and capture various poweramp states. There’s not too many controls for a poweramp (usually just resonance and presence) and there is a lot of non linearities occurring when pushing things to the limit.

Poweramp modelling is one area I think it absolutely makes sense to consider neural networks, especially as there are generally a handful of topologies that most amps use.

Presumably over time they will collect a TON of data of how everything interacts.

 

[AFKAEjay(retired)]

The one flaw I can see with Neural's approach is that it doesn't account for different speaker loads altering the behavior of the amp.

Unless they include those as a variable in the data collection…so a bunch of cabs under a switch, which you introduce as control to the user.
Seems to me…to the model it should be just another variable like any other.

 

[laxu]

I guess they could seperate capturing the preamp and poweramp (to an extent anyway), and capture various poweramp states. There’s not too many controls for a poweramp (usually just resonance and presence) and there is a lot of non linearities occurring when pushing things to the limit.

Poweramp modelling is one area I think it absolutely makes sense to consider neural networks, especially as there are generally a handful of topologies that most amps use.

I think the problem with that is that in a single amp the whole circuit will interact together and separating them might no longer work the same. Separate rack gear never seems to be quite the same thing as a single amp that shares the same transformers etc.

 

[MirrorProfiles]

I think the problem with that is that in a single amp the whole circuit will interact together and separating them might no longer work the same. Separate rack gear never seems to be quite the same thing as a single amp that shares the same transformers etc.

usually it’s the power transformer sagging the whole amp together that seems to get lost. But depending on how they measure and capture data, that could also be accounted for (possibly even modelled as an option).

The same thing could be done for a bias control and input voltage (controlling a variac).

I am interested in how they monitor and maintain that each valve is performing consistently throughout the entire time of capturing. Amps get more hot the longer they are left on, valves are always wearing out, some resistors and capacitors can change/drift

 

[laxu]

I am interested in how they monitor and maintain that each valve is performing consistently throughout the entire time of capturing. Amps get more hot the longer they are left on, valves are always wearing out, some resistors and capacitors can change/drift

My guess is that they don't. Amps can usually take it and it's not likely to make that big a difference in the data.

I used to have a Marshall JCM2000 DSL50 with that notorious bias drift issue and even that wasn't really noticeable soundwise.

 

[MirrorProfiles]

My guess is that they don't. Amps can usually take it and it's not likely to make that big a difference in the data.

I used to have a Marshall JCM2000 DSL50 with that notorious bias drift issue and even that wasn't really noticeable soundwise.

hmm, I find even a slight difference in wall voltage (which fluctuates through the day) can have a pretty noticeable effect on the tone. When the power transformer is scaling things up, a few volts deviating on input can have quite affect on the internal voltages. It’s not so much as an amp just “taking it” but how it changes its response.

Our memory for hearing and accurately distinguishing these changes may not be quite so good, but when collecting data it’s probably adding a degree of (a very specific type of) randomisation.

 

[metropolis_4]

usually it’s the power transformer sagging the whole amp together that seems to get lost. But depending on how they measure and capture data, that could also be accounted for (possibly even modelled as an option).

The same thing could be done for a bias control and input voltage (controlling a variac).

I am interested in how they monitor and maintain that each valve is performing consistently throughout the entire time of capturing. Amps get more hot the longer they are left on, valves are always wearing out, some resistors and capacitors can change/drift

That’s true there are a lot of variables at play. An amp can sound different when it’s been on a while and the tubes are heating up. Even small variances in the tubes themselves.

 

[MirrorProfiles]

I used to have a Marshall JCM2000 DSL50 with that notorious bias drift issue and even that wasn't really noticeable soundwise.

an amp with runaway bias would be VERY noticeable to hear that you’re in trouble. The giveaway would be that it goes from sounding glorious to sounding like audible mushy diarrhoea.

 

[laxu]

an amp with runaway bias would be VERY noticeable to hear that you’re in trouble. The giveaway would be that it goes from sounding glorious to sounding like audible mushy diarrhoea.

For me it never got to that point, but the point was that the bias drifting, without knowing about the issue, was not something I noticed and that would have been far more drastic than minor component variation from temperature changes.

The bias would need to go quite high, or quite low before it starts sounding terrible.

 

[MirrorProfiles]

For me it never got to that point, but the point was that the bias drifting, without knowing about the issue, was not something I noticed and that would have been far more drastic than minor component variation from temperature changes.

this can be quite context dependent - it could drift slowly over time where you just constantly need to keep lowering the bias, or it could run away very suddenly.

The bias would need to go quite high, or quite low before it starts sounding terrible.

yep, assuming the valves can take it (they all measure different out the factory). Amp settings and the amp design itself can also factor in. But there’s a reason why Fractal units have all these controls, they absolutely have an effect. And as with most things like this, they can be subtle or hard to perceive or extreme. Which is why being aware of what to expect, and also how to monitor and measure it would be important - if someone assumes they just generally behave themselves and things like the bias drifting won’t have an effect, because they couldn’t notice it on one amp, could lead to a whole stash of bad data.

 

[holoholo]

Reading all this makes me wonder if I can play guitar without being a scientist

 

[deadpool_25]

Day 2 with QC:

Messed with a WDW setup using the FR-12 in the middle and a PXM-12 on either side. It sounds great, however I think I prefer the FM9 for this kind of thing. Mainly because I can run all my post effects in parallel which just sounds special. Additionally the vast array of delays in the Fractal are , especially the SDE-3000 block that @2112 so kindly shared in his block library. The delays are fine in the QC and generally get the job done though. I wonder if any of the plug-in delays kinda get that thing going

Speaking of sharing the block library type stuff, being able to download a tone on your phone and having it instantly available on the QC is convenient AF. Man if I could have an FM9/QC mashup, with the Fractal grid and effects, and the usability and form factor of the QC...and with it being Fractal and not NDSP...

 

[BenIfin]

So if I have this right .... the TINA process Captures a large number of static Captures across a large range of interrelated Amp Control setting movements - say X - and then the ML process interpolates between them to best estimate the Amp controls in all the in-between non Captured positions across all knobs and all knob interactions to get from X to the fully captured full amp settings and full interrelated settings ... all done with zero human tweaking, refining, subjective ear-fixing etc....

Do I understand this right (?)

Ben

 

[Achilles]

subjective ear-fixing

 

[mikah912]

Day 2 with QC:

Messed with a WDW setup using the FR-12 in the middle and a PXM-12 on either side. It sounds great, however I think I prefer the FM9 for this kind of thing. Mainly because I can run all my post effects in parallel which just sounds special. Additionally the vast array of delays in the Fractal are , especially the SDE-3000 block that @2112 so kindly shared in his block library. The delays are fine in the QC and generally get the job done though. I wonder if any of the plug-in delays kinda get that thing going

Speaking of sharing the block library type stuff, being able to download a tone on your phone and having it instantly available on the QC is convenient AF. Man if I could have an FM9/QC mashup, with the Fractal grid and effects, and the usability and form factor of the QC...and with it being Fractal and not NDSP...

View attachment 25898

Fractal usability would go a LONG WAY if we could store user block libraries on the device. Not sure why that hasn't been implemented....