Most AI music companies start with a model and go looking for a problem. Just 4 Noise started with a broken drum machine.
On a grey February morning in Berlin, Max Shafer did what he always did in winter — layered up, made tea, sat down to work. He powered on his drum machine and nothing happened. So he went digging instead: 45 minutes in his sample folders, then another 15 spent staring at a disorganized hard drive, contemplating his life choices. An hour gone, no music made.
“That was the day that changed everything,” he says.
What strikes me about that story isn’t the frustration — every producer has lived it. It’s what he concluded from it. Not I need better folders. Not I need a new drum machine. Something more precise:
“I know what sound I am looking for, and if my drum machine worked, I could get that sound in just a few seconds. So shouldn’t there be a way to use machine learning to build a more robust way to search through my sample library and find those sounds quicker?”
He took the idea to Henning Nobmann, an engineer with a decade of machine learning deployment behind him and now his co-founder. Henning’s response is the reason there’s a company. Rather than build a better search, he suggested flipping the problem inside out: don’t find the sound, generate it.
“And the rest is history,” Max says.
It’s not a sample vending machine
Here’s where I think most coverage of Just 4 Noise gets it wrong, mine included until I understood what they’d actually built. “AI sample generator” makes you picture a text box that spits out a WAV file. That is not the product.
“We build multi-sample instrument generators, in every sense of the phrase.”
The distinction matters. When you generate a sound in Just 4 Noise, you don’t get a static one-shot — you get an instrument. It’s playable across a keyboard, with pitch and dynamics behaving the way they would on hardware, and it arrives with the controls you’d expect from a synth: envelope shaping, EQ, a sub bass module, multiple saturation and distortion types, and — the part I find most interesting — control surfaces derived directly from the machine learning pipeline itself.
That last detail is the tell. These aren’t generic effects bolted onto a generated file to make it feel more like a plugin — the model’s own parameters become part of the instrument. It’s closer to a synthesizer with a different front end than to anything in the sample-pack economy.
The uniformity problem
What survives when any sound is one prompt away? Producers have spent thirty years building a craft around digging, and some of that inefficiency is exactly where the happy accidents live.
His answer reframed the whole thing as an evolution rather than a rupture. First people ripped tapes, vinyl and CDs. Then came digital libraries — Apple Loops and its descendants. Now producers pull sounds from every corner of the world at once: Splice and LoopCloud on one end, independent pack creators in the middle, field recordings made by hand on the other.
Each era solved access. And each one, he argues, created the same side effect:
“This evolution has led to creators expecting to have access to all possible sounds in just a few clicks, while simultaneously creating a stream of uniformity, where every other producer uses the same exact sample or loop.”
This is the tension I don’t think the big platforms can resolve on their own, and it’s the strongest strategic case for what Just 4 Noise is doing. Their scale is the thing that flattens the output. A pack that ships to two million subscribers cannot also make you sound like nobody else. The new generation of creators wants Splice-level convenience and differentiation, and those two demands pull in opposite directions inside a library model.
Generation resolves it, at least in principle: every sound is unique on arrival, then sculpted to fit the track. Max doesn’t oversell it, which I appreciated:
“While I think this trend of generating rather than digging removes some of the happy accidents we love to reminisce on, this new workflow leaves much more room for unique exploration that larger platforms can no longer provide.”
That’s an honest trade, stated as one. I’ll take that over a press release any day.
A month in recording studios
Regular readers know that where the training audio comes from is, for me, the question that separates AI music companies worth writing about from the rest. I’ve argued before that attribution, licensing and compensation have to be built in from day one rather than bolted on afterward. So I asked.
The answer is the best one I’ve gotten from anybody in this category.
“We have a company-wide belief that you need to treat your ‘data’ with love. If you want to build models that actually sound good, you need real-world audio examples that also sound good.”
They tried the shortcuts first — forking open-source audio generation libraries, pulling down large open-source catalogs — and found quickly that none of it would support what they wanted to build. So they did it the hard way. Roughly one month in recording studios across Europe and North America, recording real analog and digital instruments to build their first training set. Then a second month running that material through post-processing pipelines and clustering algorithms to expand and label it in house.
Two months of work before there was a product to show anyone. Max’s read on it:
“While this process was super time-consuming, I still believe it brought us to where we are today, and you can really hear the difference when you use our product.”
And this is the part the industry should pay attention to: the ethical path and the quality path turned out to be the same path. Not a compliance cost, not a marketing position — a technical advantage. Going forward, as they expand their model offerings, they’re turning to enterprise customers to bring their own fully-owned internal audio catalogs to the table for training. Max expects that to be the main route from here.
Permissioned data as the default, because it produces the better model. That’s the system I keep saying we need, built by a team that arrived at it for reasons of craft rather than fear of litigation.
Narrow, on purpose
Just 4 Noise raised $1m earlier this year from BADideas.fund, SoundInvest and Sound Hub Denmark. Berlin-based, small team, going up against labs with far more compute and capital. I asked what bet they’re making that the others aren’t.
“While many other companies hunt for the consumer and try to build all-in-one AI platforms and products, we are focusing on building one thing very well: HiFi, highly controllable, multi-sample instrument generators.”
Refusing to be an everything platform is an underrated strategy in this market, and right now it’s a lonely one. Every well-funded AI music company is racing toward the same all-in-one consumer product, which means they’re racing toward the same uniformity problem the sample libraries already have.
Max’s three-year picture isn’t a destination app — it’s integration into every major platform and product offering globally. Infrastructure, not another tab in your browser. He’s blunt that the future is hybrid:
“The future of creativity will certainly be a hybrid process of automation and original ideas.”
I think he’s right, and I think the companies that survive this era will be the ones that picked a side of that hybrid and got extremely good at it. Just 4 Noise picked the side where a producer still makes the decisions — the tool generates the raw material, the human sculpts it into something that belongs in the track.
It started because a drum machine broke on a cold morning in Berlin. Two years on, it’s one of the few AI audio products I’ve used that feels like it was built by somebody who actually makes music. That shouldn’t be rare. It is.
Just 4 Noise is at just4noise.com. Max Shafer’s full origin story is on their blog.
