Licensed media training data. Cleared by the people who made it.
A generative media model is a licensing question before it is a modeling question. eeko's data practice sources music, footage, photography, illustration, scripts, and voice recordings from a creator catalogue that opted in, commissions whatever the catalogue lacks, and ships each work with its permission, its licence terms, and a signed record attached.
What is licensed media training data?
Licensed media training data is creative work, such as recordings, footage, still images, illustration, and text, that a model may learn from because the person who owns it granted that permission in writing. The grant is the whole point. A scraped catalogue of ten million tracks teaches a model plenty, and it also hands your legal team a dispute for every rights holder who never agreed.
Studios, streaming platforms, labels, and generative media startups are building models for music generation, image and video synthesis, captioning, dubbing, and scene understanding. The engineering is tractable. The provenance of the training set is what keeps a product from launching, and it is the question investors, distributors, and counsel ask first.
eeko builds production AI systems for regulated enterprises, and the data practice exists because those systems need material nobody has published: consented, human-made, checked by a professional in the medium, and carrying a record on every item. For media that means a registered creator catalogue, commissioned work to your spec, and the verification record used across the whole AI training data practice. See verified human data for the checks and AI data licensing for the terms.
Media and entertainment AI training data, by medium.
Six lines of supply, one record format. Each starts from a written spec, runs through a pilot batch, and scales only after you have checked the sample.
- Mixes, stems, and single-instrument takes
- Genre, tempo, key, and structure labels from musicians
- Exclusive grants available on commissioned sessions
- Narrative, documentary, sports, and stock
- Shot type, camera movement, and scene labels from editors
- Talent releases bound per clip
- Editorial, commercial, fine art, and illustration
- Dense captions written by visual artists
- C2PA manifests where the format supports them
- Prose, verse, screenplays, and treatments
- Commissioned writing to a style brief
- Structure and tone labels from working writers
- Script translation with a second reviewer
- Line-aligned voice recordings
- Voice-use consent recorded per session
- Label source recorded on every item
- Disputed calls sent to a second expert
- Your taxonomy or ours
What travels with every track, clip, still, and script.
A creative dataset is only as defensible as its intake. This is what gets attached to each work before it reaches you.
The provenance record, item by item
- Fingerprint — acoustic for audio, frame hashes for video, a crop-tolerant perceptual hash for images, a text fingerprint for writing, checked against the whole catalogue for duplicates and competing claims.
- Origin declaration — human-made, AI-assisted, or AI-generated, as declared by the creator and tested by generator-signature detection and C2PA and IPTC inspection.
- Consent and licence — the creator's training permission, the licence terms they accepted, and any talent or voice release, recorded per item.
- Signed provenance — a cryptographically signed record of who registered the work, when, and on what terms, with an independent timestamp.
- Expert review — a professional in the medium checks the sample, and every label records whether a creator, a model suggestion, or an expert produced it.
- Audit trail and receipt — a hash-chained log of every action on the item, and a signed delivery receipt naming each work and its hash.
Your counsel reads the receipt and the per-item terms. Your engineers load the shards. Nobody takes our word for anything.
Scraped catalogues versus licensed ones.
Enormous, cheap, and unowned. Rights holders surface after launch. The same track appears forty times under different names. A growing share was generated by last year's models, so the new one learns their tics.
- Exposed No permission, no licence, no defence
- Duplicated Repeats and conflicting claims skew the set
- Contaminated Synthetic media passes as human craft
- Unreleased Voices and faces used without consent
Smaller, specified, and owned. Every creator registered their own work through license your work for AI, set the uses they allow, and is paid each time a work is licensed. Fingerprints catch repeats, origin checks keep synthetic work labeled as such, and releases cover every voice and face.
- Permissioned Training grant and licence terms on each item
- Deduplicated Four fingerprint types, catalogue-wide
- Human-verified Declared origin, signature scans, expert review
- Released Voice and likeness consent recorded per item
From a content brief to delivered shards.
Fixed scope and a fixed price on every order. The pilot batch is priced separately, so you never commit to volume on faith.
Solutions and experts behind media datasets.
Neighbouring industries we source for.
Licensed media training data questions.
Is it legal to train AI on music, video, and images?
That question is being argued in courts and legislatures right now, and the answer differs by country and by use. The safe position is the one that does not depend on the outcome: train on work whose creator granted permission in writing. Every work eeko delivers carries that permission and the licence terms the creator accepted, bound to the item itself. We supply the documentation and the record; whether a given use is lawful in your jurisdiction is a question for your counsel, and we do not give legal advice.
How do you know a track or an image is human-made and not AI-generated?
Three checks, applied to every work at intake. The creator declares whether the work is human-made, AI-assisted, or AI-generated, and that declaration is stored on the item. We scan the file for generator signatures and for C2PA and IPTC fields that mark synthetic or edited material. A working professional in the medium then reviews a sample of each batch, because a trained ear or eye still catches what a scanner misses. Works that fail any check are held, not shipped.
What does a media training data licence cover, and can it be exclusive?
Each work carries its own licence terms: the permitted uses (training, evaluation, fine-tuning), the territory, the term, and whether the grant is exclusive or shared. Catalogue works are usually licensed non-exclusively. Commissioned recordings, footage, or writing made to your specification can be delivered under an exclusive licence, and some catalogue creators offer exclusivity for a higher fee. The terms travel with the item in the delivery metadata, so your legal team reviews the actual grant rather than a summary.
How much does licensed media training data cost?
Price follows the medium, the licence scope, exclusivity, the depth of annotation, and the volume. A multitrack session with stems and a musician's structural labels costs more per item than a captioned still. An exclusive grant costs more than a shared one. Larger volumes and standing orders lower the per-item price. We quote a fixed figure once the written spec is agreed, and the pilot batch is priced on its own so you can judge quality before committing to volume.
What file formats do you deliver media datasets in?
Audio, video, and image collections ship as WebDataset shards, which load straight into most training pipelines. Image captions and region labels can be delivered in COCO format. Labels, licence terms, and provenance metadata ship as JSON Lines or Parquet, with CSV available for tabular exports. Every dataset is described by a Croissant metadata file so it can be catalogued and loaded without custom code. Licensed image deliveries also carry a forensic delivery mark that ties the files to your receipt.
Do voice performers and people on camera consent to AI training use?
Yes, and the consent is recorded per item. Recordings of a performer's voice and footage or photographs showing an identifiable person are only offered for training when a release for that use is on file, and the release is referenced in the item's provenance record. Dubbing sessions are recorded by performers who agreed to voice use at the time of the session. If your model needs voice or likeness data under narrower terms, we write those terms into the spec and commission to it.
Tell us what your model needs to hear, see, or read.
Send a short brief: medium, volume, licence, and the model it feeds. A data lead replies inside one business day with questions, a sourcing plan, and a price for the pilot batch. No deck. A ranked list of what we can source and what we cannot.
Experienced within
Markets served.
As an enterprise AI agency, eeko systems delivers production AI systems remote-first across the United States and internationally — including these markets: