Stable Audio switched from a licensed library to Creative Commons

Stability AI's two audio models used different training-data bases and different commercial-use ceilings.

Historical event

September 13, 2023

First source published

September 13, 2023

Site publication

September 18, 2026

What happened

Stability AI launched Stable Audio on 13 September 2023 as a commercial product generating audio from 'descriptive text prompts' supplied by the user, with a free tier capped at 45 seconds and a paid tier producing 90-second tracks 'downloadable for commercial projects'. Just under a year later, on 22 July 2024, the company published Stable Audio Open, alongside a paper describing a smaller, openly weighted model built on a different training set from its commercial predecessor.

What the documents say

The 2023 announcement states Stable Audio was trained using music and metadata from AudioSparx, a music library, under a direct partnership between Stability AI and that library, a licensed-catalogue provenance claim, not an open-web crawl. Stable Audio Open, by contrast, states it was trained on 'approximately 500,000 recordings under Creative Commons licenses', broken down as 472,618 recordings from Freesound and 13,874 from the Free Music Archive, with Stability AI stating it used Audible Magic's detection system 'to ensure the removal of potential copyrighted music' from that set. The two announcements describe different rights bases for training data from the same company within one year, and the paper confirms the open model's weights sit on Hugging Face under Stability AI's Community Licence, described as free for non-commercial use and for commercial users under $1 million in annual revenue, with enterprise terms available above that.

Why it matters for makers

The mechanism worth separating is licensed-catalogue training versus Creative Commons training, since they carry different guarantees. A model trained on a library the vendor licensed directly, as the original Stable Audio was, puts the clearance burden on the vendor's contract with that library. A model trained on Creative Commons audio puts a different burden on the vendor: correctly identifying licence terms across hundreds of thousands of individually contributed recordings, which is why the announcement names a copyright filter as a mitigation rather than a formality. For a maker, the commercial-use ceiling on Stable Audio Open, free only below $1 million in revenue, is a concrete number to check against a project's scale.

What to check before you use it

Check which Stable Audio a claim refers to, since the commercial 2023 product and the 2024 open model have different training data, licences and generation-length limits, and a review of one does not describe the other. Check the Community Licence's revenue threshold against the actual scale of a commercial project, not just whether a use is commercial in kind. This is an editorial checklist drawn from the two announcements, not a warranty either one gives.

• Which Stable Audio version, commercial or Open, does a given claim, review or output actually describe?
• Does the stated training-data source match a licensed catalogue, a Creative Commons pool, or something undisclosed?
• Does the Community Licence's revenue threshold cover the actual scale of the intended commercial use?

Stable Audio's two launches, ten months apart, are a rare same-company case study in how differently 'trained on licensed data' and 'trained on Creative Commons data with a copyright filter' can be dressed up to look similar while carrying different obligations for the person using the output.

Sources & reading trail

Announces Stable Audio, its AudioSparx-licensed training data, and free/pro tier length and commercial-use limits.

Source published: 13 September 2023 · Retrieved: 16 September 2026

Announces the Creative Commons-trained open model, its Freesound/FMA data sources, Audible Magic filtering, and Community Licence revenue threshold.

Source published: 22 July 2024 · Retrieved: 16 September 2026

Paper describing the architecture, benchmark performance and the CC-BY-4.0-licensed paper itself, with weights on Hugging Face.

Source published: 19 July 2024 · Retrieved: 16 September 2026

Papers, terms and official documents establish the record; the maker reading and the checks are Signal to Song editorial analysis. This retrospective draft does not imply the site published on the event date.

Continue reading

• Meta open-sourced MusicGen with a licensed dataset and an NC licence
• Suno assigns song ownership only on its paid tiers
• Open models show two ways to disclose training data
• Browse the complete the archive

Published September 18, 2026, not on the date of the event described.