We’ve seen guitar pedals being used by synthesizer enthusiasts. So here is another candidate which might be worth your while to convert into a rack module.
From github:
NeuralPi is a guitar pedal using neural networks to emulate real amps and pedals on a Raspberry Pi 4. The NeuralPi software is a VST3 plugin built with JUCE, which can be run as a normal audio plugin or cross-compiled to run on the Raspberry Pi 4 with Elk Audio OS. NeuralPi is intended as a bare-bones plugin to build on. The pedal runs high quality amp/pedal models on an economical DIY setup, costing around $120 for hardware to build yourself.
Furthermore have a look at a video demo on YouTube:
Whenever I see made with neural networks I’m always a bit sus that it’s just marketing. Really, at the end of the day, this is a raspberry pi 4 acting as a vst host.
That’s cool in and if itself, heck my current project is a headerless raspberry pi soundfont synth.
Neural nets are often just buzzwords. If it is, building a neural net would be a good way to reduce the complexity of an emulation so that something like a raspberry pi could run it.
I may have missed it, but does the article go into what the alg does? If it works like you feed a clean signal and the transformed signal, and produces a model that will act as a novel transformer to any audio, than this is pretty nifty indeed.
-edit-
Just to be clear, im very interested in this, just a bit fatigued with the types of articles that include ML buzz for the clicks you know?
Within limits that is exactly what it is described to be. The training of the network takes some time but apparently it then transforms audio in real time at a 44100 Hz sampling rate.
Let me know when there’s a neural net that takes clean signals and a bunch of transformed signals, and produces a model for a new and novel pedal that doesn’t sound like anything already on the market
My issue with this is once you have a learned model, it might be useful for that specific effect, but exposing the hyperparameters in a way would take a lot more effort than i think would be useful.
Think about the case of a chorus effect. Once you have a model, you would likely need to fuck about with it a lot just to discover what would be the equivalent of the “rate”, and even if you did, its VERY likely that hyperparamer is dependent on a lot of others, so youd end up with a mess.
Now that i think about it, if what you’d want is a very odd mess, you’d do worse than making a really weird ML appox of what reverb does to a source sound. I bet you can get some really wild sounds that you cant get on the market. (whether or not you LIKE that sound is another matter)
Only just spotted this post.
Some very interesting ideas. Fx seems a basic use though. I wonder if you could use it for room conditioning, speaker tweaks, noise reduction etc.
There are a few AI harmonizers and percussion projects out there which could add a lot to a modular.
Ah well, another one for the project pile.
I did see the schematics however, so i guess if the code is still available it should be ok.
I’d like the pcb but a stripboard version would be acceptable as well
I havent tried this yet, but one of the projects of many ive taken on is a voice synthesis option via Tacotron 2. Not sure how portable the learned models are to NeuralPi, but that would be a cool step.
I know very little about programming but my friend does.
i’m looking for some cheap time based effects that can be synced via midi or trigger and our guess was that it shouldn’t be too hard to implement if we already had the code for a delay
