Which one of these approaches are followed by QC, does anyone know?
White-box modeling: It means that every analog component is meticulously measured and digitally simulated, so that the model not only produces the sounds of the modeled amplifier, but also allows adjusting the controls like on the modeled amplifier.
Black-box modeling: refers to profiling or capturing the sound of an amplifier with the controls set at specific values.
I’m of course referring to the MODELS and not the CAPTURES.
Well, a lot but not “all” combinations. As I understand, the algorithm running on TINA selects the most important combinations.
Its definitely a black-box machine learning modelling.
This. You simply can’t do all knob positions since they are continuous and the resulting state space is of infinite size. The continuous problem gets discretised to be solvable, therefore it’s ‘only’ an approximation (a very good one).
“A few thousand control positions”. So not all combinations. But way more than a few.
Aside from that,
I’m an electrical engineer who has done SPICE modeling. In order to have an accurate model, you need to have a precise schematic–something that most amp manufacturers are reluctant to hand over.
SPICE modeling also requires that you have correct models of each individual component in the circuit you’re building. If the amp’s potentiometers have a particular resistance curve, or any other component (caps, transformers, tubes, etc.) have unique characteristics other than what SPICE has for its base set, then those have to be modeled individually.
While SPICE modeling is important and useful–I have used it to run simulations for power electronics devices in harsh space environments–it’s also tedious as hell and takes months to perform accurately.
Black-box modeling of our beloved tube amps is the way to go.
Can this be a problem?
A Spice models can be absolute, and if i have , in years, more Power, i can open even more the computing and improve accuracy. Or with the same Power…i can do some adjustements with patches that improve the models.
With black box? Can they upgrade the captures or they have to replicate all the entire process with more accuracy? (Actually some High gain models are so so…meh)
From a user’s perspective it doesn’t really make a difference. In each case you would get an upgraded model.
Either in the form of updated modeled hardware components or in the form of a new neural network trained with better accuracy.
From an economic and engineering perspective I agree with @DiffractionCircuit.
The black box approach is much less time consuming (at least if you have something like TINA) and it has the huge advantage that you don’t have to know every detail of the hardware you are modeling.
At start yes, i agree…but in refinements?.
White box then It’s flexible, with Black box you have to re engage all Tina and assembly process…probably.
White box can be supplied with nnetw in dark points, where the model computing Is too much heavy.
And WB can be impossible to process at its perfection due computing Power limitation.
But this can be and advantage. You do the work only One time , and refine with time or Power improvements.
Consider that two triodes in spices Today with a 7800x3d with 3 sec of wav (DI guitar) in distortion can take half hour if you improve the Resolution of the model…
I think that, if there isn’ t a math Op that can improve a done and freezed captures model, black box can make the situation a bit slave of the time of capture (and state of the used neural network at Moment of capture)
Fair point, you cannot tweak a black box directly, only by doing more training.
But I think in practice we’re approaching a point where both methods are so good that the details you’d want to refine are virtually indistinguishable for most people.
So there will be less and less need for any refinement. And in this scenario, I’d prefer the method that lets you quickly train on a new amp and get it right on the first shot.
I big difference to me is that black box modeling can only reproduce amps that already exist, while white box modeling can produce new designs in the digital domain that would be impossible or impractical to physically build. The Litigator model in Helix is a good example.
That’s definitely one of the most important differences. It all depends on what you want to achieve.
In terms of sound and feel, both methods are valid. Both can be done well or not.
While there are many valid points in there, there’s also very little facts to back up claims like “No black-box approach can do that.”
If there isn’t one today, there might be one tomorrow. It all depends on the right training methods and architecture of the black box.
I think the most helpful comment in there is number 4
[4] I don’t claim our modeling to be perfect, nothing is, but at this point it is highly accurate and I myself routinely fail a blind A/B test between the models and the reference amp into the reference speaker. More importantly, though, it simply sounds good and getting caught up in “it doesn’t sound exactly the same as MY amp” is counterproductive. Make music, life is too short.
Getting back to the topic of white box or black box modelling on QC, here is an interview with Doug from NDSP where he talks about the cost and necessary knowledge involved in both methods.
After three years i’ m selling QC, because i’m more pleased from White box approach in fact. On my ears i have reached this conclusion, personal (to be clear) conclusion.
But i can undestand that this can be completely subjective.
Play Guitar and Happy new year to all!
This forum Is a good Place for discussion, and in fact i have to say, poor of fanboys.
Cheers
An AI-based capture/profile can do a better job of describing the amp’s behavior, from the user’s perspective, than component-level models, IMHO. Additionally, the captures require less processing time.