Consent and copyright in speech training data Separate recording copyright, embedded works, individual permissions and personal-data processing before licensing speech for AI. Includes a rights-review matrix. Copyright, individual voice permissions, personal-data processing and a buyer’s contract are related but separate questions. A podcast producer’s signature does not automatically answer every one of them, and public access does not itself establish a training license. Conversely, it is inaccurate to say that every use in every jurisdiction requires the same form of consent: applicable exceptions, contracts and facts matter. The practical objective is a reviewable source-specific account of who can authorize which proposed uses, what remains excluded, and what evidence supports the conclusion. Review focus Proposed uses and the actual source Separate layers Recording, embedded works, people and contract Evidence Documents tied to selected material No blanket claim Public access and a vendor label are insufficient Why one signature may leave several questions unanswered A speech recording may involve a producer, one or more performers, an employer, a publisher and owners of inserted material. These are possible contributors to investigate, not a presumption that every episode has the same ownership arrangement. A supplier needs to explain the basis on which it proposes to grant the requested rights. The [U.S. Copyright Office’s Circular 56](https://www.copyright.gov/circs/circ56.pdf) distinguishes a sound recording from underlying creative works. An interview file can therefore require a different review from a song, a scripted reading or a clip that embeds another recording. Removing a visible copyright notice or converting the audio format does not answer those source questions. Use four review lanes First identify the recording and any underlying material. Next identify individual permissions relevant to the intended treatment of the voice. Separately assess personal-data processing. Finally read the buyer’s proposed contractual grant, exclusions and responsibilities. Treat the result as a set of decisions with supporting evidence. [GDPR](https://eur-lex.europa.eu/eli/reg/2016/679) provides a personal-data framework with several lawful bases; it is not a copyright license. An individual agreement may provide useful evidence, but its wording and the actual proposed use must be examined. Calling every document a “consent form” hides these distinctions. Illustrative review: an interview with an intro music bed This example is an authored procurement exercise, not a description of a Fiund archive or agreement. The proposed buyer wants speech recognition training and evaluation. The source is a published interview with music at the beginning and a short third-party excerpt later in the episode. Lane Evidence to examine Possible unresolved issue Recording Who captured, produced and controls the selected edition The publisher holds a distribution arrangement rather than the requested training grant Embedded works Music, scripted readings and third-party excerpts A broadcast permission may not cover the new use People and voice use Guest terms and the specific planned treatment of voices Recognition use and identifiable imitation may need different review Personal data Purpose, applicable basis, notices, recipients and retention Commercial access does not settle the privacy assessment Buyer agreement Selected files, permitted acts, restrictions and downstream access A broad AI label may omit model release or subcontractors Interpret the gaps instead of hiding them If the music permission does not cover the proposal, the project may need to exclude the affected interval, secure additional rights, or choose different material. Keep the exclusion linked to source times and dataset versions. A bare statement that “ads and music were removed” is difficult to audit when there is no record of what was examined. If a guest agreement covers publication of the interview but the planned product imitates the guest’s voice, do not silently treat those as the same purpose. Escalate the mismatch for an actual rights and contract assessment. A technically excellent segment can still be unsuitable for the proposed use. What copyright research establishes—and what it does not The [Copyright Office’s May 2025 Part 3 pre-publication report](https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf) analyzes collection, training, retrieval and outputs, including fair use and licensing. It provides a framework; it is not a court decision resolving every form of speech-model training. The legal assessment depends on the jurisdiction and actual facts. Avoid two shortcuts: “the file is public, so any training is permitted” and “every AI use always requires this exact written consent.” A proposed license is one way to document an agreed scope, but the page should not replace legal analysis with a sales claim. Read the grant as an engineering specification The contract should make the permitted operations intelligible to the team implementing them. List training, evaluation, transformations, source redistribution, affiliate access, subcontracting, model distribution and identifiable voice generation separately where they matter. A permission to train does not tell an engineer whether source clips may be posted in a demo. Also specify the material: original or edited recordings, transcripts, metadata and future updates. Define what constitutes an accepted delivery and how corrections or exclusions propagate. Technical acceptance and legal scope should refer to the same IDs and versions; otherwise a later operational change can drift outside what was reviewed. Ask for evidence at the right level of access A buyer does not need unrestricted copies of private identity documents in an initial sourcing call. Request the available evidence and a controlled review process appropriate to the stage. Distinguish a supplier’s assertion from a document actually reviewed, and distinguish a document template from proof that it covers each selected recording. Identify the grantor and explain its authority for the selected files. Provide a source and component inventory, including known exclusions. Map each proposed use to the relevant documents and unresolved conditions. State whether any asserted rights, warranties or indemnities are actually offered for this project. Agree how a later correction, claim or scope change will be handled. How the AI Act fits into the packet The [EU AI Act](https://eur-lex.europa.eu/eli/reg/2024/1689) introduces role-specific obligations, including copyright-policy and training-content-summary duties for general-purpose model providers. These are not interchangeable with a source license or individual agreement. A useful supplier packet can support a buyer’s assessment by describing origin, processing and limitations accurately. It should not promise universal compliance or claim that every speaker has granted every possible use without current evidence. Fiund’s proposed source and commercial scope must be confirmed for the particular project. Does owning an episode file prove authority to license it? No. Possession establishes access to a file, not the full scope of rights in its recording, embedded material or contributors. Ask the supplier to explain its authority for the proposed grant. Does a training license automatically permit voice cloning? Do not assume so. Read the specific grant and restrictions and review any individual voice or identity permissions relevant to the intended product. Recognition and identifiable imitation can raise different questions. Must every rights document be public? No. Public descriptions can summarize evidence and limitations while actual agreements are handled through an appropriate restricted review. A summary should not claim that a document was reviewed when it was only requested. Research method: Fiund research desk; public sources, explicit worked-example labels and project-specific availability review. Editorial contact: hello@fiund.com. Sources Circular 56: Copyright Registration for Sound Recordings: https://www.copyright.gov/circs/circ56.pdf Copyright and Artificial Intelligence, Part 3: Generative AI Training, May 2025 pre-publication report: https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf Regulation (EU) 2016/679, GDPR: https://eur-lex.europa.eu/eli/reg/2016/679 Regulation (EU) 2024/1689, consolidated AI Act: https://eur-lex.europa.eu/eli/reg/2024/1689