Putting a model inside a plugin is easy. Your computer has power to spare. Putting one inside a pedal is a different job entirely.
That is what makes AI music hardware worth paying attention to. When a box has to run a neural network on its own processor, in real time, with no internet, the marketing falls away fast. Either it works at the speed you play, or it does not ship.
Roland, Akai, Neural DSP and IK Multimedia have all shipped something real here. Here is what each one is doing, and the trade-off none of them can avoid.
What AI Music Hardware Actually Means
Two very different things get sold under the same label, so it helps to separate them early.
Capture means the device learns one specific thing and reproduces it. Point it at an amplifier, let it listen, and it builds a model of how that amp responds. The result is fixed, and you play through it.
Transformation means the device reshapes audio as it passes through. Send a signal in, and a model turns it into something else on the way out. That one has to run continuously while you perform, which is a far harder engineering problem.
Both run into the same wall. A model that sounds better needs more computation, and a pedal has a small processor and a power budget. So every company here has had to decide what to give up.
Four Boxes Doing It Now
These are shipping products and working prototypes from established companies, and not just concept videos.
Roland Project LYDIA
Project LYDIA Phase II, Roland’s neural sampling prototype
Roland Future Design Lab has been building Project LYDIA, a neural sampling stompbox, with Neutone, a Tokyo AI audio company. The name is an anagram of DIY and AI, which tells you the spirit of it.
You load a neural model and it transforms audio in real time, with knobs for the parameters and MIDI on the back. Phase I arrived in November as a deliberately unfinished technology preview. Phase II added an LCD for picking models and preset memory, and showed publicly at Superbooth in Berlin in May 2026.
The detail worth noting is what it runs on: a Raspberry Pi 5. Roland is being open that this is research hardware rather than a finished product, and they have published the full Phase II write-up themselves.
Roland X Neutone – Project LYDIA Phase II
Akai MPC Stems
MPC Stems splitting a mix on the hardware itself
This is the one most producers will actually touch. MPC Stems splits a mixed recording into drums, bass, vocals and other parts, and since July 2024 it does that on the standalone MPC itself rather than on a connected computer.
The algorithm comes from zplane. It runs on the MPC Live series, One, One+, X, Xse, Key 61, Key 37 and Studio Mk2. MusicRadar published the timings when it landed, and they are honest about the cost: roughly 22 seconds to separate a 10-second sample on the hardware, against about 7 seconds on a desktop.
Akai sits under inMusic, which also now owns Native Instruments, as we covered in our piece on that acquisition.
Akai MPC Stems – Split Your Audio into Stems
Neural DSP Quad Cortex
Quad Cortex learns an amp with Neural Capture
The Quad Cortex is a six-core floorboard that models amps and effects, and its Neural Capture feature learns a physical amplifier by listening to it.
What makes it useful for this discussion is that Neural DSP shipped two versions and explained the difference. Neural Capture V1 runs offline, on the unit. V2 sends the job to the cloud for a higher-resolution result, which the company says catches behaviour like volume-knob cleanup, amp sag and bloom. Both are documented on the Quad Cortex product page.
Neural DSP Quad Cortex
IK Multimedia TONEX
TONEX captures a rig for the pedal to play back live
IK took the same capture idea and split it across software and hardware. You model an amp, cabinet or pedal in the TONEX Modeler application, then load the result onto a TONEX pedal for stage use.
That split is a reasonable answer to the power problem. Do the heavy work once, where there is power to spare, and let the box just play it back.
IK Multimedia TONEX
The Trade-off Nobody Escapes
Look at those four and the same decision keeps repeating. Run it on the device and accept a smaller model, or send it away and accept a dependency.
Akai names its two tiers plainly: the standalone hardware uses an efficient algorithm, while the desktop version uses the higher-quality one. Neural DSP does the same thing with V1 on the unit and V2 in the cloud. IK moves the whole modelling stage onto a computer and keeps the pedal for playback.
Nobody is hiding this, which is refreshing. It does mean the question to ask changes, though. Not “does it have AI”, but “where does the computation happen, and what happens when I am offline at a gig”.
An on-device model keeps working in a venue with no signal. A cloud feature does not. For a studio tool that may never matter. For a pedal on a stage, it matters a great deal.
Where the Expression Still Lives
Scale-locked bends: the model handles theory, your hand handles feel
Notice what none of these boxes touch. Every one of them works on tone or on material. Not one of them decides how you play.
That gap is where we have spent our time. We do not build hardware, and we are not about to pretend otherwise. What we build sits between the controller in your hands and whatever is making the sound.
Fluid Pitch is the clearest example of what sits between the controller in your hands and whatever is making the sound. It locks your pitch wheel to a chosen key and scale, so a bend cannot land outside it. Polyphonic mode bends whole chords through an MPE synth, and micro-tuning lets you set up Arabic maqams or Indian ragas rather than being stuck with equal temperament.
The theory is handled. The gesture is entirely yours, including how fast, how far and how hard. That is the same division the hardware above respects, applied to performance rather than tone.
It is also why we got the first place with it at the first MIDI Innovation Awards in 2021, alongside the hardware prototypes. If MPE is unfamiliar, our guide to MPE versus standard MIDI explains why per-note control changes what a controller can do.
Fluid Pitch was the winner of MIDI Innovation Awards 2021
What to Look For
Where does it compute?
On the device, on your computer, or on a server. Each is a legitimate choice, but only the first survives a venue with no wifi.
How long does it take?
Real-time transformation and offline processing are different products. Stem separation taking 22 seconds is fine while you are building a beat, and useless mid-performance. Check which one you are buying.
Is it prototype or product?
Roland has been unusually straight about Project LYDIA being research. Plenty of companies are less clear. A prototype is exciting and worth following, but it is not something to build a live rig around.
Frequently Asked Questions
What does AI music hardware actually do?
Two main jobs so far. Capture, where the device learns a specific amp or pedal and reproduces its behaviour, and transformation, where a model reshapes audio as it passes through in real time. Stem separation on a sampler is a third, sitting between the two.
Does AI hardware need an internet connection?
It depends on the feature. Akai’s stem separation runs on the standalone MPC. Neural DSP’s Neural Capture V1 runs on the unit while V2 uses the cloud. Check this before a gig rather than after.
Why is on-device AI slower than the same thing on a computer?
Processing power and heat. A pedal or sampler runs a small processor on a tight power budget, so makers ship a lighter model. Akai separates a 10-second sample in roughly 22 seconds on the hardware, against about 7 on desktop.
Can AI hardware change how I play, not just how I sound?
Not yet, in any shipping product. Current hardware works on tone and on audio material. Performance gesture, timing and expression still come from the player and the controller.
Is neural amp capture the same as profiling?
They aim at the same result by different routes. Both listen to a real amp and build a playable model of it. Neural capture uses a trained network, which is why makers describe how much of an amp’s dynamic behavior their version manages to hold on to.