On Machine Learning and A.I. Technology Used for Guitar Modeling

(excerpted from a recent GS post I made...)

The problem with sims to this point is that:

- Syntax - people using generic terms, or misusing terms
- Apples and oranges - comparing different technologies
- Non-linear vs. snapshot comparisons.


Making an IR that sounds identical to a *static* IR is straight forward, but also Occam's Razor invoking. "Why does it sound identical... sometimes?"
Algorithmic approaches feel better, and *seem* realistic but *not identical*. I can adjust my playing to sort of mimic the Real Thing, but "like a real amp" is not the same thing as "just like a real amp". I love Bandmasters because of the variety of timbre you can get from pick attack, dynamics; you can fake that with Certain Emulations, but it's not as subtle and controllable. I have to exaggerate and worse, think about doing it.

That's no fun.

And speaker IRs are only as good as the engineer that made them, what mic and where they put it on what speaker through what preamp at what level. A greenback sounds different cranked; if you play soft it acts completely different, sounds different. Again, there are ways around that, but it's a kludge and it requires conscious effort.

You can make it work "like a real amp", but not *just* like a real amp.

That's not fun.

All of these things I believe can be fixed, but I'm not sure if the programming talent and effort is in the right place, or guided suitably. They're all almost good, suitable for "most applications", but not exactly the same across all applications.

I think the *curation* at companies has been the most important thing until now, the success of some and the downfall of others. ML/a.i. may make it moot, but the "curation" of the training will still be a factor in the end result production-wise. The end result may be fully convincing, but not in the production style desired, which might be a new problem.

/ $.10