machine learning to capture analog tech

[Morphoice]

Ahoi!
Lately more and more plugin companies catch my eye claiming to use "machine learning" to capture analog technology (alias free) plus some other buzzwords...

Can someone elaborate on what we are looking at here? What are those machines? How is this done and can this be done in HISE? The only ever thing in this ballpark I heard of is the Kemper profiler... But now they are doing it to Distressors, 1176s etc... And usually compressor plugins all do pretty much the same with just different values and sensitivity curves for attack/release. At least if AP Mastering can be believed ;)))

[griffinboy]

@Morphoice

AP mastering is not to be believed, in his tests he has bad luck and chooses many plugins that have been half assed and are not proper analog circuit models, wheras, there are many good analog model products out there. Either that, or he downplays the differences. From discussing with him, his stance is not that analog model products are a sham anyway, his stance is that he prefers digital tools which are predictable and don't cause distortion. He admitted to creating his videos in such a way to clickbait for the algorithm / be polarizing. That's where the whole 'scam' thing came from, although he is right that many analog modelling plugins are not very good and don't match the real devices very well.

That out of the way, machine learning for analog simulation is mostly about creating digital models of individual circuit components, or small circuits. For example, you can train a neural network on the response of an EC66 triode tube and it will create an accurate digital model which runs much more efficiently than doing it mathematically by brute force, and may be more stable (so you may be able to eliminate recursive solvers by using Neural). This can be done for whole circuits too (but you run into aliasing issues). Then you can insert the AI model into your regular circuit simulation. The boon is more accuracy and less computational cost, all at once.

The other thing you can do is to try and train an AI to predict an entire device. For example you have a gigantic dataset of gain reduction from an analog compressor, for all it's knob settings. The AI can then be trained to match this dataset and produce the correct curves, learning the nuances of the device.

It's easier said than done. Many problems arise and getting the AI to be both correct and run efficiently are two huge challenges.

There are research papers on the topic just search for the word 'differentiable' alongside analog modelling, or a type of device eg compressor.

[Dan Korneff]

@Morphoice The basics are they are using ML models to analyze and replicate the behavior of analog equipment. They run various test signals through them and use the results to train a model that can predict how the device will respond to any input signal.
The Kemper Profiler uses a different approach. They capture the full response of an amp or effect at different settings. It creates a snapshot of the device and builds a profile. This is closer to convolution than ML.
The real problem you're going to run into with time-based dynamics (like a compressor with slow release) is the model will start to emphasize the non-linear behavior. The result of these ML compressors ends up sounding like saturators rather than compressors.
I'm on the same path as @griffinboy where I'm focusing on making models of specific components and inserting them between other analog models.
If you want to start tinkering with Neural models in HISE, here's a thread of my starting process:
https://forum.hise.audio/topic/8701/simple-ml-neural-network/108?_=1748359797127
@griffinboy also posted a thread about some details to the training scripts.
I haven't had time to fully dig in for the last few months, but plan to soon.