Generating music is a solved problem. You write "slow bossa nova, nylon guitar, 90 BPM" and a few seconds later you have three minutes of it. The technical part is no longer the hard part.
The hard part is this: can you use it? In a company video, in an ad, in a game you sell. And the answer doesn't come from the engine, it comes from the licences — which here are two, not one.
The music is generated by stable-audio-3-medium, from Stability AI, under the Stability
AI Community License: free below 1,000,000 USD of annual revenue, counting
affiliates. Above that figure you need a commercial licence from them.
⚠ And here is the part almost nobody mentions: that model embeds T5Gemma, Google's text encoder, which drags along its own Gemma Terms of Use. Two stacked licences, and complying with one does not release you from the other.
If you bill less than a million — which is nearly everyone reading this — you can use what you generate. If you bill more, this is not a detail to be sorted out at deployment time: it is sorted out beforehand, talking to Stability.
We are not your lawyers. We give you the facts so that someone who is can look at them in ten minutes instead of a week.
The licence requires "Powered by Stability AI" visibly. Not a footnote on your terms page: visible where it is used. It is a condition of the licence, not a courtesy.
There is a comfortable myth worth breaking here: "it's trained on public domain material".
It is not. It is 1,278,902 recordings: 806,284 licensed from AudioSparx and 472,618 from Freesound. Licensed material, which is a different — and better documented — situation than public domain, but not the same thing. If you are going to put a sentence about the origin of the material on your own page, make it true of this model.
A track is billed per generated second, and there is a detail we did not expect when we measured it: watermarking costs more than generating. Past roughly two minutes, the watermark is the bulk of the bill, not the model.
It sounds odd until you see why: Stable Audio 3 runs eight diffusion steps and is extremely fast — a six-minute piece generates in seconds — while watermarking walks the whole audio. The longer the track, the more the mark weighs.
The exact figure for your plan is in /v1/plans and in the documentation; we don't write it
here because prices change and an article doesn't find out.
Utteracurl -X POST https://api.uttera.ai/v1/audio/music \
-H "Authorization: Bearer $UTTERA_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt":"slow bossa nova, nylon guitar, brushed drums, 90 BPM","seconds":180}' \
-o track.wav
Up to six minutes and twenty seconds per track. seed to reproduce exactly the one you
liked, negative_prompt for what you don't want — "vocals", "drums" — and translate if
you'd rather describe it in Spanish: the model understands English far better, and we return
in X-Prompt the text actually used.
Paid plan required, from Startup up.
Like everything we generate. Inaudible watermark, required by Article 50(2) of the EU AI Act, impossible to switch off.
For music that has a practical reading worth understanding: the mark does not prove the track is yours — it proves a machine made it. If what you need is to establish authorship against a third party, the watermark is not the tool; registering the work is. What the mark solves is the legal obligation to declare the content as synthetic, and that it solves completely.
Anything to add or correct? Write to support@uttera.ai. If you correct us, we edit the post and credit you.