You clicked.
We warned you. Yes, it is AI, the old kind: music theory written as physics by a theoretical quantum physicist who should have been asleep. No generative model, nothing trained on anyone's recordings. Eleven paragraphs follow. Last exit: the button at the bottom.
Why
Generated music hands you a polished result. The fun was never the result.
Brumm gives you back the making: your melody, your timing, your mistakes, inside a track that sounds like Berlin at 6 am. Human-made, AI-assisted: music theory written as physics, searching for the version of your hum with the least energy. Nothing is generated.
Under the hood
The long answer.
For the curious, the nerds and the investors. Nobody has to read this to hum. Short version: a theoretical quantum physicist wrote music theory down as an energy, the app finds the minimum, and a resident who knows how a night is built runs the track. No generative model. Nothing trained on anyone's recordings.
The scorer: music theory, written as physics.
Your hum is read as notes by a pitch tracker, every ten milliseconds, in a background thread so the beat never hiccups. Then every note gets an energy. Moving it costs, stiffness times semitones. Leaving the key costs. Landing on a clash with the bass or another melody costs, measured at the pitch you hear, not the one on paper. A key change costs most of all.
The app runs an exact shortest-path search over every version of your phrase, every note kept, moved or dropped, in every candidate key, and plays the one with the least energy. Dynamic programming, harmonic tension, voice leading, modal keys with Dorian always on the ballot. The physics of a soap bubble, applied to a bad hum.
The pad is a force field: left to right, how much energy the fit may spend; up and down, whether your line or the track gives way. Several melodies are settled against each other the same way, whole lines moving, never the newest one. A sung take is cut into syllables and each one is pitch-shifted into place with a time-domain resynthesis, so the words stay yours.
The resident: a model of how a night is built.
We took electronic music apart. How a set opens. How layers stack: kick and bass, then a clap, then the offbeat hat, then it opens up, one per phrase. What a plateau is, when a break wants a build and when a drop wants to punch. That became a style profile and a planner: given the seed and what you did, it decides every bar, eight bars ahead, within change budgets, so you always know what comes next and can undo it.
Six grooves, fills, swaps instead of cuts, the bass never leaves, and your choices are held for a while before the resident touches them again. Pure, deterministic, testable: the same night from the same seed, every time. It is a rule set written by people who have stood in those rooms, not a model trained on anyone's tracks.
What comes next.
The physics stays. Next to it, small sequence models, recurrent networks, decision trees or whatever wins on the test set, trained on public-domain scores: classical music that belongs to everyone. Their job is to rank what the search finds, which of two fitting versions a human would have written. Trained on compositions, never on recordings, never on music that belongs to someone else.
Generative models that write the music for you are not on the list. Human-made, AI-assisted is the category the scene plays; generated is the one it bans. The fun was never the result.
Every number in the energy, every rule in the profile, was set by ear and is tested by numbers. If it sounds wrong to you, it is a bug, not a taste. If you can build it from this page, go ahead; the numbers are the part that took a year of listening.
More, if you must.
The other boring pages. Same warning, same exit: the app is one tap away.