Muzze

About us

Why we're building Muzze

We're musicians first. We write software for a living, and Muzze is where the two meet: a studio for people who make songs, built by people who make songs.

What made us start

In a couple of years AI music went from a curiosity to an industry, and we didn't like the shape it took. The biggest services were sued by the major labels, then signed deals with them. People who had made songs on them found downloads switched off, terms rewritten after the fact, and the models they'd learned to work with lined up for replacement. And hardly anyone in the business says what their models learned from, what a song costs to make, or why the price is what it is.

We wanted the opposite: a place where what you make is yours and stays downloadable, where you can see how the thing works and what it costs, and where the people who make music are the ones it's built for. Nobody was building that, so we are.

What we built, and what we didn't

We didn't train a music model from scratch. That takes a mountain of recordings we don't have and wouldn't take without asking. The Muzze v0.5 engine is built on an open-source music model, released under a license that allows commercial use, running on a GPU we rent.

Everything around the model is ours:

  • turning your words and a line of style into something it sings well;
  • making two takes of every render, listening to both, and putting first the one that sings more of your words and holds its key;
  • following a tune you hum or play, so a song can grow from it;
  • the section editor, the stem mixer and the mastering desk.

The model's authors say it learned from licensed music, royalty-free music and synthetic audio. We can't check that ourselves, and we'd rather say so than suggest more. What we can promise is our side of it: nothing you write, record or make on Muzze is used to train any model, ours or anyone else's.

What it costs us, and what it costs you

Other services keep this to themselves. Here are our numbers.

The GPU
An NVIDIA RTX PRO 4500 with 32 GB, rented at $0.72 an hour. It's on around the clock so your render starts the moment you press the button: about $526 a month, whether anyone renders or not.
One song
A three-minute song in two versions takes about a minute of it, which comes to about 1¢.
What you pay
$15 a month for 100 credits on Muzze Monthly, which is 15¢ a song. Or top up with 20 for $5, 50 for $10, and 150 for $25. While the beta lasts, nothing is charged at all.
Where the difference goes
The card processor keeps about 74¢ of every $15. The rest pays for the hours the GPU sits waiting for someone to press Render, then the web server, storage and email, and one day our time. With about 41 subscribers spending every credit, the GPU pays for itself. Until then, we pay for it.
Your invite
25 free songs cost us about 30¢ of GPU. It's the cheapest way we know to let you judge for yourself.

When these numbers change, this page changes with them.

What isn't good yet

  • Some styles come out better than others. In our own blind listening, synth-pop vocals were sometimes out of tune. Folk, country, hip-hop, reggaeton and rock held up best.
  • It's small: 50 places in the beta and one GPU. When several people render at once you wait in a queue, and the page tells you where you are in it.
  • A song made by a machine may not be protected by copyright everywhere, and some platforms ask you to say a song was made with AI. Muzze claims nothing in what you make, but that part is yours to check.
  • Some things are still rough. If a render is unusable, tell us why from its page and the credit comes back.

What changed lately

We're building this in the open, and this is where it shows. Tell us what to fix next from any song's page, or write to contact@muzze.ai; a person reads it.

  1. Songs in half the GPU time

    We rendered nine songs five different ways and listened to them blind, without knowing which was which. The one that scored best was also one of the quickest, so it's what every song is made with now: the sound of a three-minute song takes about 10 seconds to make, where it took 22.

  2. A calmer studio, and one that works on a phone

    No more purple and glow. Things you press are white, and blue means sound. Buttons, sliders and switches are big enough for a thumb, and every page fits a phone's screen.

  3. Your recordings live in your library

    A tune you upload or record is kept in your library, marked as yours, with every song made from it listed underneath. It's listened to once when it arrives, so each new cover of it starts sooner.

  4. Tell us what's wrong, from the song itself

    Every track has a way to say what worked and what didn't. We read all of it, and it's how we decide what to fix next. An unusable render gets its credit back.

  5. Terms in plain English

    The terms and privacy policy were rewritten to be read, with a summary at the top of each. When they change, every page tells you until you've read them.

Come and make something with us.

25 songs free with an invite. Leave your email and we'll send one when there's a place.

50 places left in the beta. Have an invite? Enter your code.