Podcast retrieval datasets: episode metadata, transcripts and relevance labels Podcast search datasets: episode metadata, transcript rights, relevance judgments, source-version control, supplier models and pilot budgets. Podcast retrieval datasets connect shows, episodes, transcripts or audio passages with the queries and relevance judgments needed to test search. A directory export alone is not a retrieval benchmark: it lacks the question set, judgment policy and source-version controls that make results interpretable. Buyers should distinguish metadata access, transcript rights, audio access and permission for model development. This guide explains the layers, public research references, provider models and pilot design for podcast discovery and search. Fiund can discuss a project-specific sourcing brief; the selected sources determine availability, permitted use, pricing and preparation. Primary task Find relevant episodes or passages Data layers Show/episode metadata, transcripts, audio and relevance judgments Benchmark reference TREC Podcasts Track Key identity issue Feed changes, duplicate releases and dynamic edits Pricing Metadata access and licensed source content are separate scopes What a podcast retrieval dataset contains A useful retrieval collection has at least a document universe, a query set and a policy for deciding relevance. The documents may be shows, episodes or passages within an episode. Those choices determine what an answer means. A system that finds the correct show has not necessarily located the specific discussion a listener requested. Metadata and transcripts provide different evidence. Titles and descriptions summarize what a creator chose to advertise; a transcript reveals spoken content but can be noisy or incomplete. Audio may add spoken emphasis or enable direct listening. Keep the layers distinct so a benchmark can test whether an improvement comes from better ranking, additional source content or changes in document coverage. Research evidence and actual access requests The [TREC Podcasts Track](https://trecpodcasts.github.io/) is a primary example of podcast retrieval and summarization evaluation. It establishes that these are defined research tasks, not merely generic uses imagined for a catalog. A public [Spotify research-dataset access question](https://www.reddit.com/r/MachineLearning/comments/1k5zvbf) illustrates the practical difficulty of obtaining historical material. Those sources support a useful buyer guide, but they do not prove commercial purchasing volume or guarantee access to the original corpus today. Treat a public request as a qualitative need and a benchmark as evidence of a research design. Avoid presenting either as a current procurement commitment from the organization named in the source. Major references: scale is not a substitute for task definition The [Spotify 100 K publication](https://aclanthology.org/2020.coling-main.519/) describes a large research collection of episodes, speech and associated text. [SPoRC](https://huggingface.co/datasets/blitt/SPoRC) provides transformed research material at still broader episode coverage. Their purposes, release versions, content layers and permissions differ. The [SPoRC paper](https://aclanthology.org/2025.acl-long.1222/) specifically distinguishes transformed material from redistribution of original audio. That is an important boundary for anyone comparing a transcript corpus with a commercial audio proposal. A large row count can describe metadata or derived features without establishing a matching number of licensed recordings. Choose the retrieval target first A show-level query asks which series covers a topic. An episode-level query seeks a release. A passage-level query asks where a particular idea is discussed. Known-item search, topical exploration and question answering require different judgments. A collection can support several tasks only if the evaluation keeps their targets separate. Target Example user intent Required ground truth Show retrieval Find an interview series about a field Show-level scope and relevance policy Episode retrieval Find a discussion of a particular event Episode identity and topic evidence Passage retrieval Locate the explanation of a specific concept Time-bounded supporting regions Known-item recovery Find an episode partly remembered by a listener Verified target plus plausible distractors Answer grounding Return a supported answer with a listenable source Answer evidence and source offsets A metadata schema that survives real archives Use stable internal identifiers and preserve the source identifiers as fields rather than using a mutable title as a key. Feed URLs, episode enclosure URLs and platform IDs can change or refer to related editions. Retain the observation date and source version so later updates do not silently rewrite the benchmark. Useful fields can include show ID, episode ID, release date, language, creator description, duration, source URL, content hash and archive relationship. Decide which fields are permitted for redistribution and which remain internal review metadata. A scraped directory can be useful for discovery without automatically authorizing the buyer to redistribute every description or train on the audio. Transcripts require their own provenance An episode transcript should identify the audio version, transcription process and review status. Automatically generated text can miss names, numbers or overlapping speakers; editorial text can remove the very passage a query needs to retrieve. A query derived from a transcript error can create a benchmark artifact rather than a meaningful search problem. Preserve original text and any normalized search layer separately. Record sentence or word offsets if passage retrieval depends on them. If advertisements or other regions are removed, decide whether time references map to the public release or the prepared file. An accurate search result should not send a listener to the wrong point in a different edit. Build queries from realistic information needs A useful query set includes paraphrases, partially remembered details and domain terminology, not only exact titles copied from metadata. Separate queries generated from document text from those collected independently. If a model sees the phrasing used to construct its target, results may overstate the system’s ability to handle actual users. Include answerable, ambiguous and unanswerable queries where the product needs those behaviors. Specify how evaluators should treat broad questions that match many episodes. Do not force a single relevant target when several documents legitimately satisfy the information need. Query provenance and judgment rules belong in the delivery alongside the content. Relevance judgments need a rubric and review A graded rubric can distinguish a brief mention from a sustained, useful explanation. Define what counts as directly responsive, tangential or irrelevant, and decide whether the unit is the whole episode or a particular passage. The rubric should make it possible for another reviewer to understand why a result was accepted. Disagreement is informative. Retain independent judgments and adjudication for difficult queries rather than hiding variation behind one final label. If judging only a candidate pool, document which retrieval systems contributed candidates and how unjudged material is handled. A high score on a narrowly pooled set is not proof that the entire archive has been exhaustively assessed. Versioning, feeds and dynamic insertions The same episode identity can point to changed audio over time because of replacement uploads or dynamic advertising. A retrieval benchmark should define whether it evaluates a captured snapshot or the current public feed. Keep hashes, retrieval dates and edit relationships where permitted. Syndicated copies, clips and compilations can also create apparent duplicates. Preserve relationships between source sessions and published items when available. Removing every similar title can discard legitimate editions; keeping every encoding as an independent document can distort ranking results. Describe the deduplication policy and retain the decisions in a versioned manifest. Compare providers by the layer they actually license A podcast directory or API provider may sell metadata access with rate or redistribution limits. A publisher may license selected recordings and descriptions. A transcript provider may deliver derived text under a different agreement. A research maintainer can supply a bounded benchmark with research conditions. These are different procurement routes. Ask whether the proposal includes storage, refreshes, redistribution, model training, evaluation and display of source excerpts. Do not infer audio rights from an API subscription or a transcript download. If a sourcing intermediary helps assemble the collection, the actual source permissions and exclusions still need to be established for the selected material. Price scopes and a planning example No defensible public average for a fully licensed podcast retrieval corpus was established in this research. A metadata subscription and an archive-plus-judgments project should not be compared as if they were the same product. Request separate lines for source content, transcript preparation, relevance labeling, update cadence and delivery work. For a hypothetical 100-episode pilot, suppose a team sets invented placeholders ofUSD 1,000 for scoped content access, USD 1,500 for transcript/identity preparation, USD 1,200 for query creation andUSD 1,800 for independent relevance review. The planning total isUSD 5,500. Those are transparent assumptions, not observed marketplace prices or a Fiund quote; replace them with the actual source proposal and annotation plan. A pilot should test discovery and answer grounding Select episodes from several shows, topics and recording conditions, then include plausible distractors rather than an artificially easy document set. Review a query that matches metadata well, one supported only inside the conversation and one with no defensible answer. These cases reveal which source layers actually improve the task. Check that each accepted result has a stable document identity and usable evidence span. Verify the cited timestamps against the selected audio version. Inspect whether a result depends on a removed advertisement, a transcript hallucination or a title that no longer matches the episode. The pilot report should distinguish file validity, label quality and task performance. Split and contamination controls Keep related episodes, clips and repeated source material together where the experiment requires independence. A show-held-out split tests generalization differently from a time-held-out split. Decide whether queries about later events should be excluded from training and preserve the cutoff in the benchmark documentation. If queries are generated from a language model, document that method and consider whether the same model family is being evaluated. Do not claim a contamination-free benchmark merely because its JSON files were created recently. The underlying public episodes may already be widely known, while new licensed source material can require separate evidence of prior availability. Deals and source access are different claims This research did not verify a largest completed podcast-retrieval dataset transaction with a disclosed price. A major podcast distribution agreement or advertising contract does not provide a defensible valuation for transcripts, metadata or AI-training rights. Describe any future transaction by its actual scope and date. Likewise, an open-source project using a corpus does not prove that its code license covers the underlying episodes. Source access, dataset terms, software licensing and model-development permission should each be checked separately. These distinctions help buyers understand why a retrieval-ready delivery can require more work than collecting public URLs. Discuss a retrieval collection with Fiund Email hello@fiund.com with the target retrieval unit, language and topic coverage, archive size, transcript requirements and desired query/judgment process. State whether the project needs audio, text, metadata or a combination, and whether ongoing refreshes or a fixed snapshot are required. Fiund can discuss candidate sourcing and a representative preparation plan. Availability, permission, price and delivery are confirmed for selected sources. Start with a high-level brief rather than transferring private listener queries, account information or publisher material before access and handling are agreed. Is a list of podcast URLs a retrieval benchmark? Not by itself. A benchmark also needs a defined document universe, queries, relevance judgments and version controls appropriate to the task. Do metadata rights cover the audio? Not automatically. Metadata access, transcript use, source audio and model-development rights can be governed separately. Should search retrieve whole episodes or passages? It depends on the product. Whole-episode discovery and passage-level answer grounding require different target definitions and judgments. Can I use a paper’s license for the dataset? A publication license governs the paper. Verify the corpus’s own access and underlying content conditions before assuming commercial use or redistribution. What does Fiund confirm before a proposal? The source layers, intended use, sample approach, preparation and available documentation are discussed for the actual collection. A category page is not a guarantee of current access to a historical research corpus. Sources Spotify100K official publication: https://research.atspotify.com/publications/100000-podcasts-a-spoken-english-document-corpus Spotify100K COLING paper: https://aclanthology.org/2020.coling-main.519/ TREC Podcasts Track official site: https://trecpodcasts.github.io/ SPoRC author-maintained dataset card: https://huggingface.co/datasets/blitt/SPoRC SPoRC original ACL paper: https://aclanthology.org/2025.acl-long.1222/ Spotify100K access request: https://www.reddit.com/r/MachineLearning/comments/1k5zvbf