In the last few years more and more plug-in developers that use some type of neural networks to emulate analog equipment have popped up. The most well known today is probably Neural DSP and judging by their, and others, success this seems to be the way forward for emulating analog equipment. There are a few developers I know of that have embraced this technology but I wanted to ask if you guys know of any others?

The ones I know that are or have used some type of neural networks like wavenet, RNN, CNN etc. are:
Neural DSP - WaveNet
Acustica Audio - ?
ChowDSP - recurrent (RNN)
Mercuriall - ?
GuitarML - WaveNet
Native Instruments - state trajectory network (STN)

There are probably a greater number of developers who uses neural networks to make AI assisted plug-ins, like iZotope, but I'm more interested in the ones that are emulating analog equipment.

 

Last edited: Jan 30, 2021

Mercuriall
Native Instruments (ICM models in Guitar Rig 6)

 

Ah, forgot about Mercuriall! Love the Euphoria plug-in. It's probably the one plug-in that can rival Neural DSPs stuff IMHO! From what I remember they started with neural networks in the AMT pedal emulation plug-in. I haven't demoed that one yet so I might have to download it!

Didn't know that NI used neural networks in GR6 though. Interesting. I must say that I didn't like it at all! Hahaha

 

Last edited: Jan 28, 2021

Yes, there was nothing extraordinary in GR6. The Supro Chicago model was cool though. There is also this guy (https://github.com/GuitarML) who has developed some open source plugins using neural networks.

 

They don't use NN/AI. They use non-linear convolution at several input levels (that important gain stage thing) and a shitload of patented interpolation methods to switch between those levels.
It really is hardware effect sampling, which is cool.

Edited: outdated, now they seem to also use AI, read below

 

Last edited: Jan 29, 2021

I guess they do use neural networks.
https://acustica-audio.com/pages/specials/deep-learning-basics

 

A rose by any other name would still smell sweet.
William Shakespeare (Romeo and Juliet)

 

You're right, they use it too, didn't know that
'June 27, 2019
From sampling gear to sampling humans. With Cola, we have introduced a first-of-its-kind in AI-based approaches to mixing.'

And
'We are not new to Artificial Intelligence technology. In fact, we have been using 'AI-based machine learning' while developing our compressors.

But Deep Learning techniques - applied to our already unique hardware sampling technology - will be a true game changer as - to our knowledge - the creative approach of an engineer to mixing has never been 'sampled' and recreated in a plugin.'

 

Wow, I didn't know. So now they use AI besides their classic non-linear sampling.
Cool, very cool.
Thanks for correcting me @Bubak . @No Avenger, thanks too and ban Bubak for correcting me anyways

Edited: just in case, just joking

 

Last edited: Jan 29, 2021

AI based plugins, then there's AI based synths, then there's AI based MIDI helpers, there's AI based DAWs, and one day only AI will only make music, because humans are too busy making AI

 

What acustica needs is a redesign of their code, keep the quality but with a lot lower cpu, can be done, or else they will eventually fall behind.

 

I don't think they can go faster. Couple years ago I went down the rabbit hole of their technology and there's no avoiding an humongous amount of computations. Perhaps even the addition of AI is to make them faster, partially.

 

That's why I wrote "I think?", because I did remember reading something about them using NN's but I was a bit unsure! I am definitely no expert in NN's so I wrote CNN because I just assumed that's what they are using

From what I understand, the AI used in Cola is more of a "mix helper" (like iZotope) rather than a process they have used to emulate analog gear. Machine learning doesn't have to mean that they have used any neural network either, but deep learning on the other hand would imply that they are using NN's. AA does seem to have the most accurate process for emulating analog gear right now at least! I'm hoping that Neural DSP, Mercuriall and others will apply their formulas for other stuff than just guitar amp's etc. so that we might get some great alternatives to Aqua and Nebula plug-ins

 

Ah, GuitarML is another one I forgot about haha! Their stuff sounds really good, but SmartPedal doesn't seem to work on my PC for some reason. SmartAmp works fine, even though it's a bit cpu intensive, but whenever I try to load up SmartPedal my DAW freezes until I close the SmartPedal GUI :/

 

Very cool project. I took a look, even asked some stuff to the developer. It's more a research project, he hasn't the resources of a company, neither does use the fastest libraries. He focuses on productivity and improving the quality.
The time for an optimization process is still far ahead.

 

AA pluggies still are based on IRs but I don't know in which way and to which amount they're using or supporting the plugins with AI or Deep Learning and exactly is the difference between them.
Sadly, AA doesn't tell a lot about it. Neither if they're using these techniques with other plugins too.

 

there is no clear definition of when a process is an ANN or not.

it is often only used a buzzword.

and it doesnt have much to do with AI, btw.

 

Acustica Audio uses "Vectorial Volterra Kernels" aka "VVK" (convolution/sampling) to create their plugins and libraries.

 

They're like Eq matching IRs (kernels) of the harmonics. The theory is nice and elegant.
To apply the first harmonic IR they have to multiply the incoming signal by itself, for the next another time,.. that's how they apply non-linear behavior using these kernels.
To emulate with absolute perfection they'd need to use all the content of the matrix containing the diagonal. I don't know how to even call them, say inter-modulation kernels.
But the diagonal corresponding to the harmonics contains almost all the juice.

 

True fact. Also, AI doesn't necessarily means deep learning