How Stable Audio's training data licensing works
October 16, 2026
Stable Audio 3.0 was trained on two sources: 806,284 recordings licensed from the production music library AudioSparx, and 472,618 Creative Commons recordings from Freesound, 1,278,902 in total. Stability AI published those counts in the model card and screened the Freesound portion to remove copyrighted material before training.
Key takeaways
The corpus splits into a commercially licensed catalog and a Creative Commons set, and the two were handled differently at every stage.
Screening of the Freesound material used the PANNs tagger to flag music, with flagged audio sent to a content detection company for verification.
Per-recording attribution is published for the Creative Commons subset. The AudioSparx catalog is identified by count and by source, not enumerated track by track.
Four separate licenses attach to a Stable Audio 3.0 deployment: the training data arrangements, the model weights, the inference code, and the T5Gemma text encoder.
What Stable Audio 3.0 was trained on
Two sources account for the entire corpus. Stability AI documents the composition in Section 3.5 of the Stable Audio 3 technical report (Evans et al., 2026) and restates the totals on the Stable Audio 3.0 Medium model card. Medium and large were trained on a combination of licensed audio from AudioSparx and Creative Commons recordings from Freesound.
The split is not even, and it is not arbitrary. AudioSparx supplies the larger share and the richer metadata. Freesound supplies breadth, particularly in sound effects and field recordings, under three Creative Commons categories.
| Source | Recordings | Arrangement | What it contains |
|---|---|---|---|
| AudioSparx | 806,284 | Licensed from the library | Music tracks, instruments and sound effects with text metadata |
| Freesound, CC-0 | 266,324 | Creative Commons | Recordings remaining after screening |
| Freesound, CC-BY | 194,840 | Creative Commons | Recordings remaining after screening |
| Freesound, CC Sampling+ | 11,454 | Creative Commons | Recordings remaining after screening |
| Total | 1,278,902 |
Figures from the Stable Audio 3.0 Medium model card and the technical report, Section 3.5, retrieved October 2026.
The AudioSparx catalog and the metadata that came with it
AudioSparx is a production music and stock audio library, in operation since 1996, carrying work from independent artists, composers and publishers. Its catalog runs past a million tracks and is licensed commercially into film, television, advertising and games. Stability licensed 806,284 of those recordings for training.
Metadata is the part that shows up later in the product. The technical report describes AudioSparx tracks as arriving with natural-language descriptions, titles, keywords and domain-specific fields including BPM, genre, moods and instruments. Training prompts were built by concatenating random subsets of those fields, sometimes with the field identifier kept in place, so a training example might read as Instruments: Guitar, Drums, Bass Guitar, Moods: Uplifting.
That construction is why prompt tags work the way they do at inference. A prompt that opens with a field label sits closer to the training distribution than a free-form sentence. Our explainer on AI music stems and audio inpainting covers the tag vocabulary in practice.
The Freesound Creative Commons subset
Three license categories make up the Freesound portion: CC-0, CC-BY and CC Sampling+. Terms for each are published by Creative Commons and by Freesound, and they differ in what they ask of downstream users.
The technical report notes that this is the same subset of Freesound audio used to train Stable Audio Open, the earlier open-weight model. Continuity there matters for anyone auditing the lineage of the family, because the Creative Commons portion of Stable Audio 3.0’s training data is not newly assembled material.
How the Creative Commons material was screened before training
Screening was applied to the Freesound set rather than assumed from the license tags. Creative Commons status is self-declared at upload on a community platform, so a CC-0 tag on a file is a statement by the uploader rather than a verification that the recording contains no third-party copyright.
Stability describes a three-stage filter. Music recordings inside the Freesound data were identified using the PANNs tagger. Audio that activated music-related tags for at least 30 seconds, at a threshold of 0.15, was flagged. Flagged audio went to a trusted content detection company to verify the absence of copyrighted material, and identified copyrighted content was removed.
The counts in the table above are post-filter. 472,618 Freesound recordings survived into training.
No equivalent screening step is described for the AudioSparx portion, which arrived through a commercial licensing arrangement with a library that clears its own catalog.
What Stability AI publishes, and what it does not
Two different things get called training data transparency, and Stable Audio 3.0 does one of them for the whole corpus and the other for half of it.
Composition is published for all 1,278,902 recordings: source, count, license category and content type, in both the model card and the technical report. Per-recording attribution is published for the Creative Commons portion, as a downloadable Freesound dataset CSV on the Stability AI attributions page.
The AudioSparx catalog is not enumerated publicly. 806,284 recordings came from a named library under license, but which 806,284 recordings were used is not disclosed. Not every specific recording used for training is individually disclosed.
Both figures still put Stable Audio 3.0 ahead of most of the category on disclosure, where the common position is no published composition at all. Precision, about which half is attributable, is worth more to a procurement reviewer than an unsubstantiated, broad claim about using “licensed” or “responsibly sourced” data.
The four licenses attached to a Stable Audio 3.0 deployment
Training data licensing is one layer of four, and teams evaluating the model routinely collapse them into a single question. The weights, the code, the text encoder and the training data each sit under separate terms.
| Layer | Terms | Where it is stated |
|---|---|---|
| Training data | AudioSparx licensing plus three Creative Commons categories | Technical report Section 3.5; model card |
| Model weights | Stability AI Community License | Model card; stability.ai/license |
| Inference and fine-tuning code | MIT | stable-audio-3 repository |
| T5Gemma text encoder | Gemma Terms of Use, including the Section 3.2 use restrictions | Model card access conditions |
The T5Gemma layer catches people out. Stable Audio 3.0 uses a frozen pre-trained T5Gemma encoder (google/t5gemma-b-b-ul2) for text conditioning, redistributed under Google's terms, and the model card requires acceptance of those terms before the weights download. A second set of use restrictions therefore travels with a self-hosted deployment, as covered in our guide to self-hosting a Stable Audio deployment.
Where training data provenance sits relative to output rights
Provenance and ownership are answered by different documents. The corpus composition sits in the technical report. Output rights sit in the Community License, and the license is specific about what it does and does not say.
On ownership, Section IV(c)(iii) reads: as between You and Stability AI, You own any outputs generated from the Models or Derivative Works, to the extent permitted by applicable law. Both qualifiers are part of the clause. The license allocates rights between two parties and conditions that allocation on what the law allows.
On warranties, the license goes the other way. Section IV(d) provides the materials and any outputs on an AS IS basis without warranties of any kind, naming title and non-infringement among the warranties disclaimed, and states that the user is solely responsible for determining the appropriateness or lawfulness of using or redistributing outputs. No indemnity runs from Stability AI to the licensee under this agreement. Section IV(c)(iv) runs an indemnity in the other direction, from the licensee to Stability AI, for claims arising out of use in violation of the agreement.
Licensed training data is therefore a documentation position rather than a contractual warranty. A broadcaster, distributor or errors and omissions carrier may require evidence of training-data provenance during review, and the published composition is what exists to hand over. Whether that satisfies any particular reviewer is their call, not ours, and copyrightability of generated work is a separate question covered in our explainer on whether AI-generated music is copyright-safe.
Two commercial conditions attach alongside. Commercial use requires registration with Stability AI. Where a licensee and its affiliates generate more than USD $1,000,000 in annual revenue, counted regardless of whether the revenue comes from the models at all, the granted licenses terminate as of that date and a new license must be requested from Stability AI, which Stability may grant at its discretion.
How other generative audio tools describe their training data
Disclosure in this category varies more than audio quality does. The table reports what each vendor or a named report has said, without reading anything further into it.
| Tool | How the training data is described | Published composition |
|---|---|---|
| Stable Audio 3.0 | Licensed and Creative Commons data, per the technical report | Counts published; CC portion attributed per recording |
| Suno | Conceded unlicensed training and argues fair use, per Forbes reporting | None published |
| Udio | Settled with Universal and Warner | None published |
| ElevenLabs Music | Licensed via Merlin and Kobalt opt-in deals | None published |
| Mubert | Hybrid AI and human sample library | None published |
| Google Lyria 3 | Not disclosed | None published |
ElevenLabs has a real licensed-training position, reached through opt-in arrangements with rights holders rather than through publication of the corpus. A fuller account of each platform's licensing and export terms sits in our comparison of Stable Audio against other AI music generators.
Frequently asked questions about Stable Audio's training data
What was Stable Audio 3.0 trained on?
Two sources. 806,284 recordings licensed from the production music library AudioSparx, and 472,618 Creative Commons recordings from Freesound, for a total of 1,278,902. The AudioSparx portion covers music tracks, instruments and sound effects with text metadata. The Freesound portion covers CC-0, CC-BY and CC Sampling+ material remaining after screening.
Is Stable Audio trained on licensed data?
Yes. The technical report states the models were trained on licensed and Creative Commons data, and describes the corpus as licensed audio from AudioSparx together with Creative Commons recordings from Freesound. The report's conclusion puts it as trained exclusively on licensed or Creative Commons data.
Does Stability AI publish a list of the recordings used for training?
Partly. Per-recording attribution for the Freesound portion is published as a downloadable CSV at info.stability.ai/attributions. The 806,284 AudioSparx recordings are identified by count and source rather than listed individually. Published composition covers the whole corpus; published attribution covers the Creative Commons half.
Was copyrighted music removed from the training set?
Stability describes removing it from the Freesound portion. Music recordings were identified with the PANNs tagger, audio activating music-related tags for at least 30 seconds at a 0.15 threshold was flagged, flagged audio went to a content detection company for verification, and identified copyrighted content was removed before training.
Were the small models trained on the same data as medium and large?
Not for the full pipeline. All small models were initially pre-trained on a mixture of AudioSparx and Freesound. For the final pre-training stage, distillation warmup and post-training, small-music used AudioSparx and small-sfx used a higher-quality Freesound subset. Stability reports that including sound effects data degraded musical coherence in the small models.
My recordings are on Freesound. Were they used?
Check the published Freesound attribution CSV at info.stability.ai/attributions, which lists the subset used. The same subset trained Stable Audio Open. Freesound uploads carry the license the uploader selected, and only CC-0, CC-BY and CC Sampling+ material was in scope.
My music is on AudioSparx. How do I find out whether it was included?
Stability publishes no per-track list for the licensed portion, so that question goes to AudioSparx. The arrangement runs between Stability AI and the library, and the library holds the catalog-level agreements with its contributors.
Does licensed training data mean I own what I generate?
Ownership comes from the license rather than from the training data. The Community License states that as between you and Stability AI, you own outputs generated from the models or derivative works, to the extent permitted by applicable law. Commercial use requires registration, and the $1M annual revenue threshold determines whether the Community License continues to apply.
Does licensed training data make generated music copyright-safe?
No document answers that, and the Community License declines to. Outputs are provided AS IS, warranties of title and non-infringement are disclaimed, and the agreement places responsibility for determining lawfulness on the user. Copyright treatment of AI-generated work differs by jurisdiction and remains under active debate. Take the question to your own advisors.
Does Stability AI indemnify me for generated audio?
Not under the Community License. The agreement disclaims warranties and runs an indemnity from the licensee to Stability AI for claims arising out of use in violation of its terms. Organizations needing contractual cover should raise it with Stability AI directly when requesting a license.
What is a Gemma license doing in a music model?
Stable Audio 3.0 uses a frozen pre-trained T5Gemma model for text conditioning. T5Gemma is redistributed under the Gemma Terms of Use, and the Medium model card requires acceptance of those terms, including the Section 3.2 use restrictions, before the weights can be downloaded.
Last updated October 16, 2026
