Back in 2019 when I started the neural amp modeler project, I had no clue that it would have this much impact. As a musician, it's been a privilege to see the project reach so many musicians and help them find tones that they love and help them make music in new ways. As we're getting to the end of 2024, I wanted to take a look back at the year and recap everything that's happened.
Commercial successes
You know what? Let's start with this because it's something special. NAM isn't a company making products; it's fundamentally a tech that can be used to build interesting products. So I want to start with a review of things that others have accomplished this year! In chronological order:
At the NAMM Show, Two notes Audio released GENOME, a software product that allows users to build and play signal chains in a standalone application or audio plugin. NAM showed up in its CODEX module, which allows users to load and play common Neural Amp Models* alongside other open source projects' own neural models.
Following their DAIM100 amp sim, Dual Rektifier is the second installment in Ugritone's NAM-based amp sim products.
For many people, the term "amp sim" conjures a sort of "bottom-up" approach where people meticulously break down the circuit in some gear and model it piece by piece to get the final result; neural amp modeling can provide an alternative approach, where one takes data from the signal chain being modeled and uses machine learning to develop the model end-to-end. The extreme of this is black-box "parametric modeling" (more on that below), but there are other ways that one can approach the problem. Ugritone worked with Vadim Taranov of VTarAmps to develop their NAM-based sim, so I'm not in on the details, but the backbone of the approach is NAM.
This year, we got a purpose-built pedal for NAM. DIMEHEAD is a German company who ported the open-source code to run on an embedded system, engineered the rest of the chassis and controls, and is now selling it as a product.
For a lot of people, this was a really important product, as a pedal-format piece of hardware is one of the most popular ways that people build their performing rigs today. By making this, DIMEHEAD gave an answer to a critical piece of the puzzle for lots of musicians.
I missed the chance to make a blog post about this when it came out, but Audio Assault came out with a modeling tech of their own. However, they helpfully included support for standard NAM models in their player:
It's really cool to see companies support these sort of "unified" loaders for the different kinds of data-driven models that now exist; and it's always great when doing this puts NAM in closer integration with the other features that players will want to reach for as they work their sound.
Artera DSP came on the scene this year and made waves with their macOS & iOS-based GigFast Lite product (finally, an easy way to use NAM on a phone!).
In addition to making NAM more accessible to tablet- and phone-based musicians, GigFast Lite also utilizes NAM's parametric modeling capability, allowing users to enjoy the full range of sounds that can be produced by some of the gear that Artera have modeled (instead of its behavior at a single set of parameters).
Finally, it's fast and easy to train NAMs! TONEZONE3000 is a third-party website developed around a subset of NAM's training code to streamline the process of training models.
Similarly to the Colab notebook I've maintained, the process is completely browser-based (after your record your gear), meaning there's no installation needed (assuming you already have a DAW you know how to use). Unlike the Colab notebook, the website takes you through the steps one-by-one, and it connects you to a high-powered GPU that will train far faster than the GPUs avaiable via Google Colab. You can also train multiple models in parallel, which can come in handy when you have multiple models you want to train.
Poly Effects was an early adopter of NAM, implementing a player for their Beebo pedal. The Ample, released this year, is a more amp-centric pedal, using NAM to provide the core of the tone for the tens of amps that are all contained on the pedal.
What I like about this pedal is that someone could look at it and not even notice that NAM is there--one can enjoy the large catalog of amps, all in a little box--and not be missing what the pedal essentially is. It emphasizes the difference between the tech underneath and how it's presented as a product to musicians.
Transistor Legacy Virtual is an audio plugin that Ghost Note have developed to give their customers an idea of what their guitar pedal products sound like so that users can try them out at home with their own gear.
I love this idea because of how it acknowledges and embraces several truths of playing guitar today: (1) guitarists want to be able to try out products in a way thatallows them to really see how they work, (2) many guitarists know how to use plugins, even if their main rig is hardware-based, and (3) guitarists still want to play hardware! Giving us a (NAM-accurate) preview of their products is a great way to conenct with customers and is something that I hope we'll see more companies embrace as we move forward.
Progress in the open-source project
Now, I want to swtich gears and look at the open-source resources that I maintain and develop. This year, I've released 16 new versions across the three main repositories: 6 new versions of NeuralAmpModelerPlugin (the plugin), 9 new versions of neural-amp-modeler (the trainer), and one new version of NeuralAmpModelerCore (the C++ DSP code for running models in real time in plugins, etc). The repositories are being actively maintained and I'm working through bugs and features as I have time. Since NAM is fundamentally an open-source project, my prioritization has been to favor work that will have the largest "force-multiplying" effect that others building with NAM will be able to leverage. [This is probably worth its own blog post.]
Next, I want to revisit some of what I think are the highlights in 2024 from my personal output:
Parametric modeling and ParametricOD
I released something new & different at the start of the year: ParametricOD, a plugin that demonstrates the use of NAM's parametric modeling abilities to give a complete model of an overdrive pedal through the full range of all of its knobs and switches.
Over the years, I've gotten a stream of countless messages and emails from musicians and NAM enthusiasts asking for a way for NAM to model the knobs on their gear so that they don't have to navigate through a bunch of "snapshot" models.
As I'd mentioned at the start of the year, this is a common idea, and, sure enough, NAM supports making these parametric models. However, there's a fair amount of subtlety in making them well, and I'm not currently in a place where I'm looking to share and support an open-source "simplified" parametric trainer as I've done for the "snapshot" case. However, parametric models are very much a thing, and you're going to see more and more of them in the future.
Analog-digital Calibration
This isn't (shouldn't be) strictly a "NAM thing", but NAM is nevertheless the first to demonstrate an implementation that I think people would be happy to see more of in the future.
In a nutshell, the question is: When you have a digital signal on your computer, how much voltage was that in the physical, analog domain? You can take the same voltage produced by your guitar's pickups and make it all sorts of amplitudes in your computer by turning the gain knob on your interface to scale the signal. But this introduces a "conversion factor" that can vary wildly; if you don't keep track of it, then your experience of a model might be very different from someone else's (or the experience of "plugging directly into" the gear).
I added the ability to add metadata to models trained with NAM's "standardized" trainers that can be used by plugins when playing them to account for that conversion factor that's "baked in" during the recording process that gives the data for training.
Like I said, this isn't a question that only NAM needs to address (any digital model based on an analog device deals with this), but as NAM is so accurate to real-world gear, it's now becoming more common for musicians to expect a perfect recreation of their experience playing the source gear; makig it quantitatively accurate requires calibration.
[Of course, you can always gain-stage your digital signal chain however you'd like! It bears repeating what the end goal is: to make music. Don't let "right and wrong" get in the way of you achieving your real goal as an artist!]
Epilogue: NAM is still getting bigger and bigger!
The project's "viral moment" about 2 years ago was something I've never experienced before, so it can be hard to tell where things are now relative to back then. By all metrics I can see, interest is still steadily growing, and more people are directly using the resources I've shared every day. Add on top of this all of the products that have been introduced by companies who are building with NAM, and things look even better.
When I initially wrote on the homepage that "NAM powers the next generation of digital audio effects," it was just something that sounded about right based on the interest I'd seen. Fast forward to now, and the number of things that have been released (and the number that have yet to be announced!) make me more convinced now of where things are headed. You're going to see NAM's tech in more and more places, doing more and more things.
See you in 2025!
Footnotes
*I need to get better at calling these "NAMs" instead of "NAM models"! "ATM Machine", anyone? [Back]