Music

Describe the music you need and get a finished piece, up to six minutes and twenty seconds long. These are not loops or fragments to assemble: it is a piece with a beginning and an end.

It is meant for anyone who needs background music with a clear licence —a video, a podcast, a presentation, a waiting room— and does not want to fight a stock catalogue or somebody else's rights.

Output is 48 kHz stereo WAV. The model generates at 44.1 kHz and we upsample to 48 before delivering; upsampling loses nothing the model did.

All audio is watermarked. It carries an inaudible watermark identifying it as machine-generated. It is required by Article 50(2) of the EU AI Act and cannot be turned off, neither over the API nor from the Studio.

Powered by Stability AI. Music is generated with Stable Audio 3 by Stability AI, trained on 1,278,902 licensed recordings from AudioSparx and Freesound. What you generate is yours, to the extent explained in your plan's usage licence.

Length and quality

There are two controls, and both change what it costs. That is the important difference from Sounds, where a clip costs the same however long it runs.

Length. From 10 seconds to 6 min 20 s. The piece is built whole for the length you ask: a three-minute piece is not a thirty-second one repeated six times.

Quality. This is the number of steps the engine takes, from 32 to 128. More steps means more rhythmic coherence and more structure, not more audio fidelity: bandwidth is the same at 32 as at 128, measured.

Below 32 steps you save nothing. It is measured: 16 and 32 steps take the same time, because at those figures the real work falls below the noise. That is why the control starts at 32 and not at 16.

And a third control that costs nothing: the seed. The same description with the same seed gives the same piece. If you like one and want to come back to it, keep the number.

Over the API

One route. Requires a paid plan, from Startup up.

POST /v1/audio/music — JSON body, returns the WAV.

Field	Type	What it is
`prompt`	string	Required. What music you want. Naming the style, the instruments and the BPM helps a lot.
`seconds`	number	Length, 1 to 380. Defaults to 60.
`steps`	integer	Steps, 32 to 128. Defaults to 32.
`seed`	integer	The same seed with the same description gives the same piece. If you do not send one, the engine picks one and returns it to you.
`negative_prompt`	string	What you do NOT want —«vocals», «drums»—. Sending it turns guidance on; leaving it out does not.
`translate`	boolean	`true` translates your description into English before generating.

The response carries these headers:

Header	What it says
`X-Seed`	The seed actually used. If you sent none, this is the only way to repeat the piece.
`X-Steps`	Steps applied.
`X-Audio-Seconds`	Seconds of audio delivered.
`X-Generations`	How many generations were run.
`X-Watermark`	Which watermark it carries.
`X-Translated`	`1` if your text was translated.

Limits and cost

The ceiling is 380 seconds. Beyond that it is not that it fails: the model does not know how, and we would rather say so than quietly deliver something worse.

Charged by length and by quality

Unlike Sounds, here the cost depends on both, because so does the engine's work:

credits = 0.333 + seconds × steps × 0.001167 + seconds × 0.02184

The last term is the watermark. Past about two minutes, watermarking costs more than generating: that is not an arithmetic slip, it is that the engine is very fast and watermarking scales with length.

A one-minute piece at 32 steps costs about 3.9 credits. The same at 128 steps, about 11. The application shows you the number before you generate, precisely so that moving a control is never a surprise.