Models and Licences — What Makes the Music in Aleatoria

Aleatoria makes music with two kinds of engine: our own synthesizer and nature engine (code we wrote, with no trained model in it) and open-weight AI models that run on our own GPUs in the EU — ACE-Step 1.5 for AI songs, Stable Audio Open Small for AI nature textures, and HTDemucs with basic-pitch for AI stems and MIDI. This page lists every model that is wired into the product, what it is used for, its licence and which plans may use it. Without an account you can hear the synth in a 30-second preview on the trial page. No external AI API is used: nothing you type or generate leaves our servers for another company's model.

Open the synth engine in the studioOpens the studio with this template filled in. Nothing is generated until you press Generate.

Why do we publish this?

Three reasons. The commercial licence of the Basic, Studio and Pro plans is only worth something if the licences of the models behind it allow commercial use, so you should be able to check that yourself. The EU AI Act (Regulation (EU) 2024/1689, Art. 50) asks providers of AI systems that generate audio to label it and be transparent about it. And DJs, labels and distributors ask which model made a track before they accept it. Everything below is written from the code and the licence files we ship with; the date at the top is when it was last checked against them.

Which models make the music?

Engine	What it does	Model and licence	Plans
The synth (Aleatoria Procedural)	Instrumental tracks with stems and MIDI, BPM and key written into the files, the same seed giving the same track	Our own code, no trained model (© Cloudpeakify s.r.o.)	Every plan, 0 credits
The nature engine (Aleatoria Nature)	Synthesised nature scenes and the nature layers of a hybrid track	Our own code, no trained model	Every plan, 0 credits
ACE-Step 1.5	AI songs with vocals and lyrics, AI instrumentals, and the AI studio tools (extend, repaint, remix, variation); Auto lyrics are written by its own language model	MIT (the ACE-Step 1.5 repository licence). In production the `acestep-v15-xl-sft` checkpoint, with `acestep-v15-turbo` baked into the worker image as the fallback; the `acestep-5Hz-lm-1.7B` language model ships inside the same model repository; the Qwen3-Embedding-0.6B text encoder is Apache-2.0 upstream	Basic, Studio and Pro: credits, 5 credits for every started 4 minutes
Stable Audio Open Small	AI nature textures, clips of up to 11 s, only when you choose it explicitly (the default nature engine is our own)	Stability AI Community License (gated weights); the "Powered by Stability AI" notice it asks for is shown wherever its output is offered	Every plan, 2 credits per clip
HTDemucs / HTDemucs-ft with demucs 4.1.0	AI stems — vocals, drums, bass and other — of a track made in your account	MIT	Studio (paid tool), Pro (included)
basic-pitch 0.4.0 (Spotify)	MIDI transcribed from the AI stems	Apache-2.0	Studio (paid tool), Pro (included)

Around the models: Rubber Band (GPL-2.0-or-later) time-stretches the DJ tools' renders as a separate process on the server, and FFmpeg encodes the MP3 and FLAC files. Neither is shipped to you. The model weights are downloaded from Hugging Face once, when the worker image is built; at run time the GPU worker can run with no internet access at all.

How is the output labelled (EU AI Act, Art. 50)?

Every exported file carries a machine-readable marking that the audio was generated by AI: the ICMT comment of a WAV, AI_GENERATED=true in a FLAC, TXXX:AI-GENERATED=true in an MP3, the X-Aleatoria-AI-Generated: true header on every download, and the same label in the interface. We label the purely synthesised tracks too, not only the AI model's: the provenance record of a track says which it was — trainedAlgorithmicMedia for output of a trained model and algorithmicMedia for the synth. That record is not yet embedded in the file or signed; a signed C2PA manifest is a voluntary addition we are preparing. The details, obligation by obligation, are in our AI Act document (in Czech) and summarised on the legal page.

The file tags also carry the tempo and key: a synth track has both, a track from the AI model carries the BPM only when it was measured reliably, and never a key tag.

Who owns what you generate?

You do, to the extent of your plan: Basic, Studio and Pro grant commercial use permanently for tracks made while the subscription was active; Free is an account without a subscription, which keeps its library but renders nothing while the service is paid-only. The exact wording, including the exceptions, is section 8 of the terms; the terms govern, not this page. We do not use your tracks or prompts to train models, we never compare them with other users' work and we never publish them (section 3 of the privacy policy).

What do we not use?

No external AI API. No OpenAI, Anthropic, ElevenLabs or any other hosted model: every model runs on our own GPUs in the EU (our list of subprocessors, section 4). The code contains an optional lyrics writer that would call the Anthropic API behind a feature flag; it is switched off in production. Switching it on would be a new subprocessor, announced 30 days ahead in that list.
MusicGen (Meta): its weights are CC-BY-NC 4.0, non-commercial. The engine registry knows it and would restrict it to Free (an account without a subscription), but it is not installed in the production worker image and production has never run it.
HeartMuLa-oss-3B (Apache-2.0): an optional sidecar in the code, not installed in the production worker image.
Stable Audio 3: evaluated in A/B experiments only; there is no production adapter for it.
YuE: not integrated and not used.
Recorded drum samples: the repository carries CC0 one-shots behind a switch that is off; the synth's drums are synthesised.

FAQ

Where does the music get rendered?

On our own GPU and CPU workers on Google Cloud in the EU (region europe-west4, the Netherlands). The model weights are fetched from Hugging Face only when the worker image is built; no user data is sent there or to any other model provider.

Can I use AI songs commercially?

On Basic, Studio and Pro, yes — the licences of ACE-Step 1.5 (MIT) and Stable Audio Open Small (Stability AI Community License, free for a company below one million US dollars of annual revenue) allow it, and the terms grant it permanently for tracks made while the subscription was active. A model with non-commercial weights (MusicGen) is refused for paid plans by the server itself, and it is not deployed anyway.

Why does the studio show "Powered by Stability AI"?

The Stability AI Community License asks for that notice wherever output of the model is offered, so it appears in the model picker, on a finished Stable Audio track and on the legal page. The ACE-Step licence (MIT) asks for nothing in the interface.

Does an AI song carry a key tag?

No. A song from ACE-Step carries the BPM in its tags only when the beat could be measured reliably, and no key tag. Tracks from the synth carry both BPM and key, with the Camelot code in the file name.

Where is the complete licence record?

The full inventory with pinned versions and the open questions marked for legal review is kept in our repository next to the code; this page is its public summary, last checked on 11 October 2026. When a model, a licence or a plan rule changes, this page changes with it.