AI Music Training Data Licensing and Provenance

A music file can be technically perfect and still be unsuitable for a training dataset.

The pipeline may need to answer several separate questions. Who controls the musical composition? Who controls the sound recording? Did the contributor have authority to provide it? Are performers or other people represented in the recording? Which licence or agreement applies? Does it permit the intended commercial, research, redistribution, or model-development use?

These questions cannot be answered from the waveform alone. They need documentation. That documentation becomes part of the dataset, just like the audio, captions, and tags.

One recording can involve several layers

Swipe sideways to view the full comparison

Layer	What it concerns	Possible evidence
Musical composition	The music and lyrics underlying the performance	Publishing agreement, composer declaration, public-domain research, or composition licence
Sound recording	The fixed recording of a particular performance	Master-use licence, label agreement, owner declaration, or recording-specific licence
Performance and contributor permission	Permission from performers or contributors where relevant	Performer release, contributor agreement, session contract, or collective agreement
Privacy, publicity, or personality interests	Rights or legal interests connected with identifiable people, names, or voices	Release, consent record, privacy assessment, or jurisdiction-specific legal basis
Dataset and annotation rights	Rights in the database, metadata, captions, labels, or compiled dataset	Dataset licence, terms of use, annotation agreement, or database-rights notice
Contract and platform restrictions	Promises or limits created by agreements and access terms	Platform terms, API contract, data supplier agreement, or access policy

The composition and recording are separate

A song and a recording of that song are not the same copyright work. The musical composition covers the underlying music and lyrics. The sound recording covers a particular fixed performance.

This means one audio file can involve different owners and permissions. A songwriter or publisher may control the composition, while an artist, producer, label, or another party may control the master recording.

A pipeline should therefore avoid one vague field called copyright. Store the composition and recording status separately.

In practice

Public-domain composition, modern recording

A composition may be in the public domain while a recent recording of it remains separately protected. A training policy must check the recording itself rather than assuming that the age of the song clears every audio file.

Questions to ask about a licence

What exactly is licensed?
Check whether the terms cover the audio, composition, metadata, annotations, dataset compilation, or only some of them.
Who granted the permission?
Record the licensor and the evidence that they control or can authorise the relevant rights.
Which uses are allowed?
Separate research, commercial training, evaluation, fine-tuning, redistribution, publication, and output use.
Which conditions apply?
Check attribution, noncommercial limits, share-alike terms, access controls, reporting duties, and other restrictions.
Where and for how long?
Record territory, start date, end date, renewal, withdrawal, and termination terms when relevant.
What is still unclear?
Do not convert uncertainty into an approval. Flag the record for review.

Creative Commons still requires careful reading

Creative Commons licences use standard elements such as attribution, noncommercial use, share alike, and no derivatives. The exact combination matters.

The relationship between copyright licences and AI training is legally complex. Creative Commons published separate guidance because copyright laws and exceptions vary between countries and uses. A conservative project may choose to follow licence conditions even where the legal requirement is uncertain.

The technical lesson is simple: store the exact licence name, version, URL, attribution information, and the policy decision made for the intended use. Do not reduce every Creative Commons record to a field that only says open.

Common Creative Commons checks

Swipe sideways to view the full comparison

Element	Plain-language question	Pipeline response
BY	What credit and source information must be preserved?	Store creator, title, licence, source URL, and required attribution text
NC	Does the intended use fall within a noncommercial restriction?	Block automatic commercial approval and request policy or legal review
SA	Could share-alike duties affect distributed adaptations, models, data, or outputs under the chosen legal approach?	Record the issue and obtain project-specific review before distribution
ND	Does the planned processing or distribution conflict with no-derivatives terms under the applicable interpretation?	Do not guess. Send the planned acts and distribution model for review
CC0 or public-domain tool	Does the tool cover the relevant asset and are third-party rights still present?	Preserve the declaration and continue checking recording, performer, privacy, and third-party material

In practice

Real dataset example: MAESTRO

The official MAESTRO page states that the dataset is available under CC BY-NC-SA 4.0. A pipeline designed for commercial model development should not treat the word `Creative Commons` as automatic approval. It should record the noncommercial and share-alike elements and apply its commercial-use policy.

A licence addresses specified rights in a work. Consent records a person's agreement to a defined activity. The two can overlap, but neither automatically replaces the other.

For example, a producer may control a recording but may not have permission to authorise every use of a guest vocalist's identity or voice. Creative Commons notes that third-party privacy, publicity, and personality rights are not automatically cleared by another person's application of a CC licence.

A creator opt-in system should therefore ask what the contributor controls, which people appear in the asset, and whether additional releases or legal bases are needed.

What a contributor or consent record may need

Swipe sideways to view the full comparison

Field	What it records	Why it matters
Contributor identity	Who gave the permission	Connects the agreement with a real person or organisation
Authority statement	Which rights or permissions the contributor claims to control	Shows whether the contributor can approve the proposed use
Covered assets	Exact files, recordings, stems, voice samples, or future uploads	Prevents one agreement from being applied too broadly
Purpose	Training, fine-tuning, evaluation, search, release, or another use	Keeps permission tied to an understandable activity
Commercial scope	Whether commercial products and paid services are included	Avoids treating research permission as commercial permission
Duration and territory	Where and for how long the permission applies	Supports expiry and regional policy checks
Withdrawal or termination	How future use changes if permission ends	Allows the technical team to apply the agreed process
Agreement evidence	Signed document, version, timestamp, and integrity hash	Makes the decision auditable

Provenance connects the evidence to the data

Data provenance records where an item came from, who or what handled it, and which activities changed it.

For music data, provenance can link a processed training segment back to the original recording, source URL, supplier, agreement, licence snapshot, ingestion event, cleaning operations, captioning process, and dataset version.

W3C PROV describes provenance through entities, activities, and agents. MLCommons Croissant can use PROV-O relationships such as wasDerivedFrom, wasGeneratedBy, and wasAttributedTo at dataset, file, record, field, or value level.

Three useful provenance relationships

Swipe sideways to view the full comparison

Relationship	Meaning	Music-data example
wasDerivedFrom	One item came from another item	A mono 32 kHz segment came from the original stereo master
wasGeneratedBy	An activity created or changed an item	The segment was generated by cleaning pipeline version 2.1
wasAttributedTo	A person, organisation, or system was responsible	The caption was written by an annotator or generated by a named model version

Why provenance starts at ingestion

Sources can change
A URL, account, licence notice, or platform page may later disappear or be updated.
Files lose context
A copied audio file does not preserve who supplied it or which agreement covered it.
Processing creates descendants
One original can produce stems, crops, embeddings, captions, and token files.
Rights decisions may change
A later policy update, dispute, withdrawal, or expiry must reach every derived record.
Audits need evidence
A team cannot reproduce a decision when the licence, contributor statement, and processing history were never stored.

Useful pipeline decisions

Swipe sideways to view the full comparison

Status	Meaning	Typical next step
Approved	Required evidence is present and the intended use matches policy	Allow the record into the specified pipeline
Restricted	The asset is usable only for stated purposes, products, or environments	Apply technical access and use controls
Review required	Evidence exists but the interpretation or authority is uncertain	Send the record and evidence to the correct reviewer
Quarantined	Required evidence is missing, conflicting, or under dispute	Keep it out of training while preserving the source and issue log
Rejected	The intended use clearly fails the project's policy	Exclude it and record the reason
Withdrawn or expired	Permission ended or the agreement changed	Apply the documented removal, retraining, or future-use process

music-rights-and-provenance-record.json · A simplified record that separates rights evidence, consent, policy decisions, and technical lineage.

Swipe sideways to view code

{
  "assetId": "recording-0042",
  "source": {
    "url": "https://example.org/assets/recording-0042.wav",
    "collectedAt": "2026-07-22T14:30:00Z",
    "supplierId": "creator-018",
    "sourceType": "creator-upload",
    "originalSha256": "7c2f...9ab1"
  },
  "composition": {
    "status": "creator-controlled",
    "rightsholders": ["creator-018"],
    "evidenceIds": ["agreement-771"]
  },
  "soundRecording": {
    "status": "creator-controlled",
    "rightsholders": ["creator-018"],
    "evidenceIds": ["agreement-771"]
  },
  "contributors": [
    {
      "contributorId": "creator-018",
      "role": "performer-producer",
      "permissionEvidenceId": "agreement-771"
    }
  ],
  "licence": {
    "type": "custom-ai-training-agreement",
    "version": "2.0",
    "commercialTraining": true,
    "redistribution": false,
    "attributionRequired": true,
    "startsAt": "2026-07-22",
    "endsAt": null
  },
  "consent": {
    "basis": "contributor-agreement",
    "scope": ["training", "evaluation", "commercial-model-use"],
    "withdrawalProcedureId": "withdrawal-policy-v2"
  },
  "provenance": {
    "wasDerivedFrom": "recording-0042-original",
    "wasGeneratedBy": ["resample-32khz-v3", "segment-30s-v2"],
    "wasAttributedTo": ["creator-018", "ingestion-service-v4"]
  },
  "policyDecision": {
    "policyVersion": "commercial-music-training-v3",
    "status": "approved",
    "decidedAt": "2026-07-22T15:10:00Z",
    "reasonCodes": ["recording-cleared", "composition-cleared", "contributor-authorised"]
  }
}

Licensing, Consent, and Provenance

The composition and recording are separate

Public-domain composition, modern recording

Questions to ask about a licence

What exactly is licensed?

Who granted the permission?

Which uses are allowed?

Which conditions apply?

Where and for how long?

What is still unclear?

Creative Commons still requires careful reading

Real dataset example: MAESTRO

Provenance connects the evidence to the data

Why provenance starts at ingestion

Sources can change

Files lose context

Processing creates descendants

Rights decisions may change

Audits need evidence

Ready for a quick check?

Ready to continue?

The composition and recording are separate

Public-domain composition, modern recording

Questions to ask about a licence

What exactly is licensed?

Who granted the permission?

Which uses are allowed?

Which conditions apply?

Where and for how long?

What is still unclear?

Creative Commons still requires careful reading

Real dataset example: MAESTRO

Consent is not the same as a copyright licence

Provenance connects the evidence to the data

Why provenance starts at ingestion

Sources can change

Files lose context

Processing creates descendants

Rights decisions may change

Audits need evidence

Review Music Dataset Rights

Ready for a quick check?

Ready to continue?