Chapter 01
Where it all began...
I'm Bene, one of the builders of ToneKeep. I will tell you the story of our plugin. How an idea distilled itself in my head over months. (Note that its just me, the other guys would probably tell you something entirely different.)
Back in September 2025 I was building fluid simulations in [mesh.], our local builderspace. Its a quite inspiring place, and you get to talk to a lot of cracked people. One curious week two guys told me about the same concept: tone matching. Basically you hear a tone on a record, and you want to acquire it with a click of a button.
I was hooked, but my intuition suggested that its impossible to execute with the quality I wanted.
Chapter 02
Different ideas
LLM integrated solutions felt flimsy, and white box modelling waaay too complex. An end-to-end ML is like super impossible (nothing wrong with that though). If we just look at the SOTA stem separation models, their output is kinda phasey, which is not a good place to start in itself.
At that time I was already considering some kind of hybrid approach. A big inspiration of mine was actually a videoseries by Jim Lil: where does tone actually come from. [link] So regarding amps its basically how you put distortion and EQ blocks after each other.
This limits the parameterspace in a quite meaningful way in my mind -- the first win.
The century old question
Where does tone come from?
Chapter 03
GP-50 and ML
In the next year I had a full month break from uni (i finished exams super early). I wanted to dig more deep into ML, my idea was to find topological defects in spacetime in public domain sky surveys. I also happened to impulse-buy the brand new GP-50, which blew my mind (I think every semi-serious guitarist needs one).
I also got to know TONE3000 and NAM, suuper nice stuff. And the connection was made immediately for me. I need to create a parametrized version, as a basis for the tone matcher. This basically solves the tone quality problem, and gives me a way to construct the ML stack for the tone matching. This felt insane, I validated the theory, with ML experts from [mesh.] in just 3 days. I was thinking about building this asap.
I started training DEMUCS modells for de-distorting guitar signals on our H100-s, and the started thinking about the forward modell on a 5090 ([mesh.] going crazy btw). Also I made a couple of demos about my ideas.
I had a great time, but some things were missing.
Chapter 04
Pivoting
First of all the tonematching thing is a little gimmicky if not executed right. For example check out [Groundhog] What do you think? To me another problem started to seem even more relevant. We have the open source static snapshots, with incredible quality, also Neural DSP plugins are awesome, but I needed a parametrizable modell for tonematching to even happen. So it became clear to me that focusing on the “forward modell” is the right call.
So the reason while people dont create neural DSP plugins of their own amps is simple. The algorithms used for getting it done are super data hungry. You have to capture a shitload of different settings, to train the proper thing. My solution is basically that we create a couple of “base modells” that can be custom fitted to your amp with a couple of data points. (Remember that the tone stack is the most important, so if that matches, the rest is fine tuning.)
It still didn't feel complete though, and I had to figure out, that a plugin cant just run my pytorch inference scripts. You need Cpp...
Chapter 05
Meet the team
Back in late august I was at a house party, where the background music was my responsibility. I did Vulfpeck -- what else. Someone asked the host guy: “is this a random playlist, or is there another vulffan here?”. Thats how I met Dávid a math student with impressive Cpp skills and a big passion for music.
With Donát we were already friends in my local builderspace (if you remember he was one of the guys telling me about tone matching). We already did some jamming together, and always found the time to geek out at the annual builder nights. Him being an electric engineer, and having deep personal interest in the fusion of music and technology, was a huge win.
Theese two built a working prototype in 1.5 weeks, while doing university. We presented at the biggest [mesh.] demo night yet, it was a banger. And Donát brought his friend: Dani, the founder of Schweigert-Fusion, an emerging Hungarian jazz-funk band. He liked the idea, and getting him on our side is a huge boost regarding outreach.
What about the other guy telling me about tone matching? He is Mete. A crazy builder and software dev with a lot of experience and great intuition in sales. This website is his creation.
With the Avengers assembled, we are ready to build.
Chapter 06
Outlook
We are now launching this beta for a mini-market-validation, and we want to build a core community and team of early adopters, who can constantly give us feedback and help us shape the future of ToneKeep.
Hopefully we can grind trough the development as quickly as possible. When one of our models gets 95% there, we will expand our scope regarding outreach, and we will go on a campaign to create the datasets for the base modells, based on the best amps we can find, and the project will become a little more public. We are expecting this step to happen early summer, if everything goes according to plan.
After this we need to perfect the product, and get ready for a launch. Our goal is by the end of summer.
Stay tuned!!
Made with physics and mild obsession.