Speech training data cost: a practical budgeting method Separate source licensing, collection, transcription and preparation costs. Compare accepted speech hours and use a dated annotation-service example without inventing market averages. A useful speech-data budget separates permission to use source material from collection, annotation, preparation, review and delivery. Public transcription prices can help estimate one work package; they are not market averages for a dataset license. Commercial scope, accepted coverage and required labels vary too much to justify one universal per-hour price. Compare proposals using the same billable unit and acceptance definition, then state which costs and uncertainties remain outside the quote. Comparable denominator Accepted underlying speech hours Separate line items Source grant, capture, labels, QA and delivery Published example Rev annotation-service pricing checked November 13, 2026 Fiund pricing Confirmed for the selected source and scope Start by defining the thing being purchased “One hour of audio” can mean a published episode including advertisements, a complete session, the time during which anybody speaks, or the summed duration of several participant tracks. These measures can all be useful, but they cannot be swapped when comparing prices. Ask how overlap, silence, failed files and multiple editions are counted. Also distinguish the transaction. A recording license grants an agreed scope of use. A transcription order produces text from material you supply. A custom collection pays for a production process. A quote may combine these, but the budget should still identify what each line buys and how it will be accepted. A dated public annotation-price example [Rev’s published pricing page](https://support.rev.com/hc/en-us/articles/18893487380365-Pricing), checked November 13, 2026, lists starting human transcription at US$1.99 per audio minute and AI services at US$0.25 per minute. It separately lists transcript timestamping at US$0.30 per audio minute. These are vendor service prices, not Fiund prices or speech-dataset licensing rates. Use a specific service, currency and quote date. Language, service selection, tax, extras and current terms should be checked with the provider. A timestamping service should not be assumed to meet a project’s word-level boundary tolerances or annotation acceptance rules merely because it carries a timestamp label. Illustrative order Arithmetic Amount before other selected charges 100 hours of human transcription at the listed starting rate 100 × 60 × US$1.99 US$11,940 Timestamping for the same 100 hours at the listed add-on rate 100 × 60 × US$0.30 US$1,800 Those two selected service lines together US$11,940 + US$1,800 US$13,740 What the example does—and does not—purchase The calculation demonstrates the annotation component of a budget. It does not establish permission to use the recordings for model training, provide source recruitment, acquire rights from participants, remove third-party material, or guarantee dataset-level speaker separation. Those need their own scope and evidence. Nor does dividing the combined amount by 100 create an “average cost of speech data.” It creates a service estimate for the stated assumptions. If the project needs corrected verbatim text, overlap labels, adjudication or a particular export schema, ask whether that work is included and price any gap explicitly. Normalize two hypothetical proposals Consider two invented proposals for a narrow pilot. Proposal A charges US$4,000 for candidate material plus US$1,200 preparation, with 80 hours ultimately accepted. Its observed cost is US$5,200 ÷ 80 = US$65 per accepted hour. Proposal B charges US$6,000 inclusive, with 100 hours accepted: US$60 per accepted hour. The lower normalized price does not automatically make B better. It might lack the needed locale, deny an intended use, or repeat a few speakers. Add coverage and scope columns beside cost. A procurement comparison should identify the proposal that resolves the target requirement, not reward the largest loosely defined duration. Why zero license price is not zero project cost [OpenSLR lists LibriSpeech](https://www.openslr.org/12) as a read-English corpus under CC BY 4.0. The [Creative Commons deed](https://creativecommons.org/licenses/by/4.0/) permits commercial sharing and adaptation subject to its conditions; describing all open speech licenses as noncommercial or universally restrictive is inaccurate. A team still needs to assess fit, comply with the actual terms, preserve attribution where required, ingest the files and run evaluation. Read audiobooks may be useful for one baseline and poorly matched to another task. That is a source-style limitation rather than a reason to invent a commercial licensing fee or a universal ban. Build a cost model that exposes uncertainty For custom work, budget collection setup, recruitment, recording, failed sessions, transcript preparation, annotation review, delivery and project coordination separately. Some costs are fixed and some vary with candidates, accepted units or review effort. Do not assume that doubling accepted hours doubles every line. For an existing archive, show the cost of reconciliation, exclusions, schema mapping and the required technical transformations. Preserve the distinction between an estimate and a reconciled inventory. If acceptance loss is unknown, compare scenarios with different usable yields and explain which pilot measurement will narrow the range. Questions that make quotes comparable A good first brief gives suppliers enough information to identify the source and work rather than guess an all-inclusive headline. The following questions help reveal excluded work before a commitment. What is the billable unit, and how are tracks, overlap, silence and duplicates counted? Which source uses, recipients and delivery editions does the price cover? Which labels, normalization rules and export fields are included? How is the representative sample selected, and what blocks acceptance? Who pays for correction, replacement, rework or unavailable coverage? Which transfer, storage, security-review, tax or future-update costs remain outside the quote? When a project-specific quote is the honest answer There is no verified Fiund-wide price schedule or market-average range established by the sources in this guide. Speaker availability, task fit and permitted scope must be reviewed for the proposed material. An exclusive grant and a limited evaluation sample are different purchases, even when both are described in audio hours. Send the intended task, coverage matrix, preferred source type, target accepted volume and label specification. A useful next step is a bounded proposal with clear assumptions, sample review and a way to revise scope when the pilot reveals missing conditions. That is more actionable than an unsupported universal rate. Is a transcription rate the same as a training-data license rate? No. Transcription is a preparation service for supplied audio. It does not by itself acquire the source or grant the uses required by a model project. Should a quote count both channels as twice the hours? Only if that is the expressly agreed billable unit. Underlying conversation duration and summed participant-track duration measure different things. State which denominator is used and apply it consistently. Can Fiund quote an exact project from a target number of hours alone? Hours alone leave source, coverage, rights and annotation work unresolved. Provide the model task, intended uses, accepted-unit definition and technical requirements so an actual proposal can be assessed. Research method: Fiund research desk; public sources, explicit worked-example labels and project-specific availability review. Editorial contact: hello@fiund.com. Sources Rev pricing support page: https://support.rev.com/hc/en-us/articles/18893487380365-Pricing LibriSpeech ASR corpus, SLR12: https://www.openslr.org/12 CC BY 4.0 deed: https://creativecommons.org/licenses/by/4.0/