Provenance in speech training data: source lineage and permission evidence Connect speech files, excerpts, transcripts and review records without treating source possession or a complete-looking manifest as blanket AI permission. Provenance explains where a speech record came from, how it changed and which evidence supports its proposed use. Technical lineage and permission review are related but separate. A checksum can identify a file; it cannot establish what a participant agreed to. Fiund reviews source fit, documentation and commercial scope for the selected project rather than asserting that every recording in every possible source is already cleared for every AI use. Two evidence layers Technical lineage and permitted-use review Public manifest Evidence references, not private forms or signatures Status vocabulary Available, reviewed, pending, unavailable or not applicable Download Constructed provenance and evidence ledger Start with the recording, not a generic consent badge A buyer needs to understand what the proposed source actually is: a published episode, an original session, a participant track or a processed excerpt. These may contain different material and have different relevant parties. The fact that one party can supply a published mix does not describe every component of the original recording or every intended downstream use. Write a source statement that can be checked. Identify the responsible source, recording context, date range, retained representation and known exclusions. Mark estimates as estimates when an archive has not been reconciled. Avoid saying all material has a complete chain of consent until the selected material and the supporting evidence have actually been reviewed. The purpose of a provenance record is to make review possible. It should expose unresolved questions rather than compress them into a reassuring but unexplained cleared flag. A source with a pending document may still be worth evaluating, but its status should remain pending until the relevant review is complete. Technical lineage: entities, transformations and versions The [W3C PROV-O recommendation](https://www.w3.org/TR/prov-o/) provides a vocabulary for entities, activities and agents and for relationships between them. A speech preparation workflow can apply that distinction to original recordings, segmentation, transcript generation and reviewed annotation versions. The standard models provenance; it does not certify the legality or accuracy of a particular record. Assign stable identifiers to original sources and their derivatives. Record hashes, duration, channels, encoding and the processing version where those facts have been inspected. An excerpt needs its source identifier and boundaries; a transcript needs the recording or segment it describes and its creation method. A corrected transcript can change version without changing the audio hash. Preserve relationships through conversion. A WAV derived from an MP3 is not evidence of an original uncompressed capture. A source-separated participant stream remains a processed derivative, even if its filename says speaker1. Keep the actual process visible so the buyer can decide whether that representation supports the intended model task. Permission review: scope, evidence and unresolved questions Document the proposed use precisely enough for the responsible reviewers to evaluate it. Recognition training, evaluation, speech generation, a named voice and downstream dataset redistribution can raise different source questions. A permission statement should be read in its actual context rather than summarized as AI allowed without qualification. Identify which parties and components require review. A podcast may include hosts, guests, producers, music, advertisements or licensed inserts. A recording source may have a relevant agreement but lack a document for a particular component or use. This guide supplies a project-review structure; it does not give a legal conclusion for a specific archive. Keep private evidence in an appropriately controlled location. A delivery manifest can reference an agreement or review item without containing signatures, private contact information or full participant forms. An authorized reviewer should be able to resolve the reference and see the status, date, scope and any restrictions that affect the proposed delivery. Claim Technical evidence Separate review question This excerpt derives from a source Source ID, offset and checksum relationship May that excerpt be used as proposed? This transcript describes this audio Annotation version and aligned reference What uses of the transcript are permitted? A source supplied the file Transfer and responsible-party record What authority and restrictions apply? A review item is complete Dated decision and evidence reference Does the decision cover this exact scope? Public access and research release terms are not a blanket offer A paper, dataset card or public audio link can help a researcher understand a source. It does not establish an unrelated supplier's ability to resell the material. Review the selected release, its access procedure and its terms. Keep the source statement attributed when it has not been independently verified for a proposed delivery. The [Spotify podcast corpus publication](https://research.atspotify.com/publications/100000-podcasts-a-spoken-english-document-corpus) is a research reference for a particular corpus design. The [SPoRC card](https://huggingface.co/datasets/blitt/SPoRC) has its own access and research-use scope. Neither should be relabeled as a Fiund commercial catalog merely because the underlying content is associated with podcasts. Dataset documentation such as [Datasheets for Datasets](https://arxiv.org/abs/1803.09010) helps record composition, collection and intended use. A completed document is useful evidence, but it is not a substitute for source-specific permission and suitability review. Keep the distinction visible in proposals and public pages so a buyer can tell what has been observed, stated by a source or still needs confirmation. A worked evidence ledger that keeps pending items pending Download the [illustrative speech provenance ledger](/downloads/speech-provenance-evidence-ledger.json). It contains a fictional source recording, a derived15-second segment, a transcript relationship and six pending review items. There are no real participant records or agreements in the download, and its placeholder hash is not an inspected file checksum. The example deliberately separates entities from review questions. A valid segment offset describes a technical relationship even when the proposed use remains unresolved. Conversely, a relevant agreement does not prove that a delivered file matches the selected source or that a transcript corresponds to the right version. Both layers need checks. In a real ledger, add an evidence reference and dated decision only when the review actually occurred. Preserve a history of changed scope. If the buyer later requests a new model use or downstream sharing arrangement, reopen the affected review item rather than assuming that a previous decision covers an expanded project. Inspect the chain at a pilot before extrapolating Select a pilot that includes representative source families and known complications. Verify that source IDs resolve, hashes identify the delivered versions, excerpt intervals fit the source duration and annotations point to the expected audio. Check a mixture of ordinary and difficult records instead of only a curated showcase. Then review the supporting source and permission documentation through the agreed channel. Record which files or groups the evidence covers, what remains unknown and what restrictions affect preparation or delivery. An evidence record that refers to a show title alone may be insufficient to resolve alternate edits, excluded episodes or third-party inserts. Keep a reconciliation report between the pilot's proposed and accepted inventory. If an interval is removed, update its derivative records, duration totals and any dependent transcript or label references. A deleted segment should not remain counted in a release total or silently survive in a second export under a different filename. Version changes, corrections and withdrawal handling A release should state how corrections are delivered. A transcript fix, source replacement, metadata correction and scope restriction are different kinds of change. Give them distinct reasons and identify which downstream records are affected. A new archive filename alone is not an adequate change log. Where an agreement specifies removal or restriction handling, record the operational steps that follow from it. The website should not promise universal deletion from every downstream model or copy without an actual agreed mechanism. Describe what can be traced and changed in the selected delivery, and escalate questions about broader effects to the responsible project reviewers. Retain enough history to explain a prior decision without republishing removed material or private documents. Use stable references and controlled access. The goal is a reviewable release history, not a public archive of participant information. A public technical manifest should reveal relationships and limitations while respecting the handling agreed for the underlying evidence. What Fiund can review for a proposed speech project Send a high-level description of the source and the intended model use, plus the recording level, approximate volume and required annotation layers. Identify which documents exist and which questions remain open. Begin with a requirements summary rather than uploading recordings, participant forms or private agreements into an ordinary email thread. Fiund can use that information to review fit, source relationships, preparation and the scope of a proposal. Actual source availability, evidence access, permitted use, acceptance and commercial terms are confirmed for the selected project. This replaces an unconditional catalog-wide permission claim with a process that can be checked against concrete material. For a source-package example, use the Podcast Datasets [session-to-data-package walkthrough](https://podcastdatasets.com/resources/podcast-session-data-package/). For task-specific derived labels, read [speech instruction-tuning datasets](/data-types/speech-instruction-tuning-datasets/). Keep the underlying source, annotation and review evidence connected as the project moves from a pilot to a larger release. Does a checksum prove that a recording is licensed? No. It identifies file bytes. Permission depends on the relevant source evidence, proposed use and agreement. Should consent forms be placed in a public manifest? No. Use evidence references and appropriate private review access rather than publishing participant forms or signatures. Is every possible Fiund source already cleared for every AI use? This site makes no such blanket claim. Availability, documentation and permitted use are reviewed for the selected collection and project. Research method: Fiund research desk; public sources, explicit worked-example labels and project-specific availability review. Editorial contact: hello@fiund.com. Sources PROV-O: entities, activities and provenance relationships: https://www.w3.org/TR/prov-o/ Datasheets for Datasets: https://arxiv.org/abs/1803.09010 100,000 Podcasts: A Spoken English Document Corpus: https://research.atspotify.com/publications/100000-podcasts-a-spoken-english-document-corpus SPoRC author-maintained dataset card: https://huggingface.co/datasets/blitt/SPoRC