Great video, thank you for sharing! No, GuitarML and Neural are not affiliated, its just a fun open source project I work on in my spare time. Lauri was kind enough to help when I got stuck on making WaveNetVa and pedalnet compatible with each other (already existing inference and training repos). I appreciated the section on the difficulty of writing the inference code from scratch, because I just went though that process on my next plugin for GuitarML. I started with Keras, then converted that to python/numpy, then once I got that working, c++ with the help of some other open source libraries. I tried some of the Keras model to c++ converters, but just not fast enough for live audio.
I cannot begin to tell you just how much I enjoyed this video--not just because I work in AI and deep learning, but also because back in high school I was the lead guitarist of a metal band. Greak work in creating and producing this video!
Glad to hear you talk about audio and DSP. Maybe you can make videos sharing some of the interesting papers you're reading. Doesn't need to be big in quantity but just great in quality, just like the videos you have shared. There are some other guys already doing that. I think they're not as good at explaining things.
Around minute 20:00 he describes how Line6/Fractal etc are currently doing analog modeling, and he talks about filters and wave shapers. Well, I do not agree at 100% on that, mainly because that is an "emulation" approach, while today we have many options for "simulation". In other words, if we know the circuit, we can simulate it in real time building non-linear transfer functions from the combination of the characteristic equation of each component (for example the resistor is described by V=RI). There are many examples of this approach, with different algorithms, someone uses Wave Digital Filters, others use MNA (modified nodal analysis) etc... Of course in order to run in real time with low CPU usage these approaches are aided by the use of tabulated functions and other approximations.
Great video, really appreciate the insight into the different approaches to modeling. I'm hoping we well eventually evolve to a hybrid approach that allows ML to deal with some of the hard things to model, but allow the use of the parameterized blocks to build new kinds of amplifiers for new sounds that don't exist in the analog world.
How in this finite world am I gonna get that Explorer ??? And why didn't I know an Explorer could be so incredible beautiful ? Who in my family can I sell to get it?
Thanks for the video, really great content! Do you know how they solve the parameter scanning of the captured device (not spice), is it done by a DSP post-processing step? More credits to your playing in the intro, it was tasty! Congrats!
One of the writers at @7.09 is named "Eero-Pekka Damskägg" - Damskägg literally means "Ladybeard" in Swedish. I thought that was funny. Now, the first name (Eero-Pekka) seems Finnish so it might have a different meaning, but funny still.
Yeah! Thank you for that content! I am a hobby musician drummer making noise with guitares and physicist to earn a living. That's so cool. Besides I learn some nice stuff that I can come up with in the job. :) I do medical physics in a radiotherapie department. And we use deep learning to contour the organs at risk in the CT scans. But you are not talking about neural networks with a radiooncologist. Basically you're lucky if they know the power button of the computer. Haha, Sorry!
When did Neural DSP claim that their plugins are based n pre-trained neural net models? They always go on as modelling specific equipment, and by that I mean traditional modelling. The QC is supposed to be the unit that incorporates neural nets. Is there some proof/evidence that the pluns contain any of the QC technology?
@@welcomeaioverlords this is far from what claimed by you in the video above. Nolly are out over a year ago. Except for Gojira all other plugins are also relatively old and have been advertised as mkdelling platforms. Furthermore, the article above does not claim the use of neural nets for those models. Rather, they just mention the umbrella term black box which many companies do anyways one of them being Kemper albeit not using neural nets. So my question remains since in the video you explicitly imply the use of neural nets for those plugins.
@@giominor88 Neural DSP has not detailed the tech they use at the product level to my knowledge. That said, I'm not sure why a company would call itself "Neural" DSP, make several early hires of leading experts in using ML for blackbox modeling of guitar sounds and then *not* use it in their tech stack, but it's entirely possible. It never occurred to me to ask.
