IK and the Amplitube line really sucks in this regard..

[Scotty]

Delete double.

[Scotty]

That’s good to know. You’ve been speaking in generalities and haven’t revealed your qualifications previously. Kudos. Can you explain why AI modelling of the domain s covered by VIR is computationally less intensive than IR type technology. That would be helpful. This doesn’t seem intuitive to me. How would a computer scientist estimate the compute power needed to build an interactive model that allows for solutions for cabinet microphone and room modelling? I looked for academic papers on microphone modelling with AI and didn’t turn up anything let alone the interactions of various cabinets , speakers within rooms. It would seem tremendously expensive computationally. If I am wrong what am I missing?

I won’t argue over what you feel might happen. That’s fine. I would welcome someone with real computer science and machine learning qualifications to weigh in on the topic. Sometimes people talking sh*t doesn’t cut it.

I have a computer science degree and I’ve implemented Apriori, Perceptron, Naïve Bayes, and Decision Tree classification algorithms with sklearn in Python.

And I have also solved convolution integrals by hand.

[jamcat]

It takes a lot of power to train the ML model. But the result is actually a pretty simple weighted statistical model that multiplies the input by a series of weights assigned during the training. Traditional amp modeling uses layers of differential equations that involve solving ancillary equations and integrating factors and lots of integral calculus. Convolution itself is the integral of the product of two functions (the input signal and the impulse). It's all really ugly calculus, whether you're a person or a computer.

I actually really hate math and I'm not too keen on coding either.

Anyways, you might notice that a TONEX amp+cab uses about half the CPU power that an AmpliTube 5 amp+cab does.

[Scotty]

Interesting. So do the sheer number of statistical weightings increase significantly if you were to represent complex interactions between domains re: mics, rooms cabinets. Would this still be reasonably economical on a modern processor or is there no way to know for certain until you build the training models and run them? I understand hat tonex uses less CPU than amplitube but it a a snapshot of a specific settings and limited otherwise. Amplitube is continuously variable. When we are talking about interactive AI models with swappable speakers, mics and variable room sizes the statistical weighting would need to have a lot more combinations and permutations if I am understanding correctly.

On a personal note are you working professionally in the coding field now or are you making your music your paying gig or some of both? We’ve interacted on several threads and I’ve not got a sense of this.

It takes a lot of power to train the ML model. But the result is actually a pretty simple weighted statistical model that multiplies the input by a series of weights assigned during the training. Traditional amp modeling uses layers of differential equations that involve solving ancillary equations and integrating factors and lots of integral calculus. Convolution itself is the integral of the product of two functions (the input signal and the impulse). It's all really ugly calculus, whether you're a person or a computer.

I actually really hate math and I'm not too keen on coding either.

Anyways, you might notice that TONEX uses about half the CPU power that AmpliTube 5 does.

[jamcat]

Any given setting in AmpliTube is a "snapshot" too. It's not like AmpliTube is having to calculate every amp setting at once at all times. The variables within the tube model differential equation will change depending on plate current, screen voltage, control grid voltage, etc. But only one value per variable is being computed at any given time.

For machine learning, the interactions of the mic and cab and room are already baked into the ML model. I don't think a more complex system would increase the power needed to run the model, but outputting a really accurate model of such a system might take longer for the model to converge during training. However, stacking multiple isolated ML modules would take more CPU power to run. AmpliTube already stacks several processor modules for stomps, amps, cabs, and racks. So separating TONEX into separate stomps, amps, and cab captures wouldn't be any more taxing than it already is in AmpliTube.

What I think would work really well is capturing each component in isolation with machine learning, i.e., the amps without stomps or cabs, then making separate ML captures of the cabs across a 30 x 5 grid with an SM57 and U87, on and off axis, or whatever VIR uses. Then simply replace each IR with its ML counterpart. Each mic position "snapshot" is analogous to a single IR. It would require redoing all of the IR captures they did with training data instead (which they might not be too keen on doing.) But it would absolutely sound so much better, and give users the same miking flexibility as VIR. They probably could even use machine learning to swap mics with far better results than the current deconvolution+convolution.

Currently I'm working on some development ideas independently and producing my own music.

[donkey tugger]

Who invented piano, violin and guitar?
do re mi fa so la ti do - Italian!
Who was the first successful super star with a big concert tour? Ever heard of
Paganini?
What roots do Frank Sinatra and Zappa have?
Who is called the best guitarist in the world? Ever heard of Al di Meola?
Is it coincidence that Elvis looked very Italian?
Ruffalo, Cavallo and Magnoli - ever heard these names who were managers of one of the biggest success in music history?
The productions of the most successful
German producer Diether Bonham is based on Italo Disco.
Which country is the country of opera?
In Italy plugins were already known before the first plugin was ever released.
They weren't called plugins.
As a kid my neighbor Giovanni let me watch his small computer screen in a small dark room.
On the screen there were black and white tracks and I could take a listen sometimes on headphones only, even though he had speakers and he always asked me not to say anything about his music productions.
It was in 1980s. Now he's not alive any more. The music generally would not be what it is without Italian inventors, developers and artists. Italy is the mother country of music.

Yet they haven't yet managed to perfect SUPER plugin.

[acousticglue]

What some people feel will be a downer, others do not. Simple.

[Bombadil]

I've said this one before, but it's worth repeating. Non-musicians have not, and will not be able to tell the difference. It's only 'us' looking for the Holy Grail of Virtual Tone. If the song is shite, no amount of realism will save it.

[donkey tugger]

Can't remember who said it, but, "There's no such thing as a bad guitar sound, you just haven't found the right song for it", makes a lot of sense to me.

[Burillo]

Can't remember who said it, but, "There's no such thing as a bad guitar sound, you just haven't found the right song for it", makes a lot of sense to me.

plenty of songs with shit guitar sounds...

[VitaminD]

Did IKM ever get their download app to also authorize/register software in the same app?

For the longest time they had a kludgey system of custom shop app for downloading and some separate service manager app for authorizing it. Which was super annoying.