Commercial speech datasets compared: sources, preparation and pricing

Select a source family before comparing prices

Catalog conversation, custom recording, podcast archives and public benchmarks solve different problems. A catalog can be quick to evaluate when its existing languages and capture conditions fit. Custom collection can target missing conditions but adds recruitment and production work. An archive can provide natural long-form content while requiring more source, edit and permission review.

Public benchmarks are valuable for baseline experiments, yet their accessibility does not make them a complete commercial procurement solution. Read the selected release's license, composition and intended-use limitations. The cost of downloading a file is not the total cost of preparing a training release, and a free download does not resolve every permission question for a different use.

Write the requirement that would make an option fail. Examples include absent original tracks, unknown speaker reuse, a noncommercial-only release, missing transcript conventions or an inability to preserve continuous timing. Remove incompatible options before doing per-hour arithmetic. A low number on an unsuitable source is not a saving.

Two dated catalog price references, with their actual units

On 13 November 2026, SpeechData.ai’s own price page states a USD60–95 list range per hour of recorded two-person conversation, varying by language. The described bundle includes dual-channel audio, aligned human-verified transcripts, metadata, commercial licensing and consent documentation. This is the vendor's statement, not an independent audit of its inventory or evidence.

The oto full-duplex buying guide describes a different specific offer: otoSpeech 1,125-hour packs of 100–1,125 hours at USD10–20 per hour. That guide also gives a broader USD10–100 licensed-catalog range and says custom collection is quoted by scope. Do not replace the pack-specific rate with the broader range or assume every label layer is included.

Neither source establishes the price of a Fiund project. Confirm the exact selected dataset, minimum order, accepted-hour definition, transcript coverage, permitted use and any extra labeling before procurement. Prices can change; retain the date and URL in the purchasing record.

Normalize hours without inflating the recording supply

One hour of two-person conversation may be one shared-session hour and two participant-track hours. If only 42 minutes contain accepted speech under the agreed rule, the accepted speech total differs again. A supplier should disclose which quantity is billed and whether silence, overlap and rejected intervals remain in that denominator.

Calculate a comparison with an explicit formula: total project fee divided by accepted unique source hours, alongside the supplier's original unit. Keep one-time setup, recurring access and optional annotation separate. If the rights fee depends on use or duration, a single blended rate can hide the condition that matters to the buyer.

A constructed example illustrates the issue: a 100-hour session package with two isolated tracks is not 200 hours of independent conversation. If ten hours are rejected, a USD5,000 hypothetical package has a USD55.56 effective cost per accepted session-hour, before any additional preparation. This arithmetic is a planning example, not a quote, observed transaction or estimate of prevailing prices.

Use task-fit tests for ASR, generation and interaction

ASR projects need an agreed transcript convention, language coverage and evaluation split. Speech generation may require different signal quality, speaker permission scope and expressive coverage. Full-duplex interaction needs a preserved timeline and usable overlap evidence. Avoid declaring a collection suitable for all three because it contains clean speech.

The Multilingual LibriSpeech release is an example of audiobook-derived read speech, not a catalog of spontaneous support calls. The FLEURS card provides a multilingual research reference with its own recording and evaluation design. Compare those properties with the intended deployment rather than quoting corpus scale as a substitute for fit.

Ask the supplier for a deliberately mixed pilot: routine material, edge conditions and known limitations. Review the same sample across audio, transcripts and metadata. A demo assembled from only the cleanest clips can answer whether the pipeline works at all but cannot establish the rejection rate for a bulk archive.

Transcription labor and rights access are different purchases

A transcript service can transform audio a customer already controls. That fee does not normally purchase an unrelated recording archive. A data-access agreement can permit selected recordings without including human transcription or task labels. Keep those lines separate so the buyer can identify which work is actually being quoted.

If a project requests word timing, overlap labels, speaker grouping and translations, specify the convention and acceptance test for every layer. Automatic output, human correction and specialist adjudication have different scopes. A phrase such as fully annotated should be expanded into fields, examples and review criteria before comparing suppliers.

Do not infer ownership or broad clearance from technical possession. Ask which party can authorize the proposed recording, transcript and derivative uses, and how evidence will be reviewed. A public sample can demonstrate formatting while carrying a narrower access scope than the proposed commercial delivery.

Compare documentation and delivery, not only media files

Request a source manifest, version identifier, checksum inventory, metadata dictionary and change log. For multiple sources, retain source-level group identifiers and known duplication relationships. Ask how rejected or replaced items affect the release version and accepted totals.

Review permission evidence through an appropriate private channel. An evidence reference can appear in a manifest without publishing the underlying participant form or contract. Track available, pending, not applicable and unavailable separately. A document marked pending should not be summarized as reviewed or cleared.

Dataset documentation practices such as Datasheets for Datasets support this kind of explicit record. They do not certify legal suitability or performance. A complete-looking form still needs source-specific review, and a claim made by a provider should remain attributed until verified for the selected delivery.

Use the RFQ worksheet to expose unknowns

Download the speech RFQ comparison CSV. It lists source families and two dated vendor-price references with deliberately unknown scope fields. It is an editable comparison structure, not a recommendation to purchase a particular provider or a claim that Fiund resells those collections.

Add the model task, language/locale, unique speakers, session duration, capture setup, transcript convention, review sample and allowed use for each shortlisted option. Record the source URL and date when a price changes. Keep the original supplier unit even after computing an effective accepted-hour comparison.

The worksheet should lead to a short list of questions, not an artificial ranking from missing data. An unknown isolation status is a review item, not a zero score. Mark an option ineligible only when it actually fails a required condition. Retain the reason so a later source or specification change can be assessed without restarting the entire comparison.

A pilot-to-purchase process that preserves the evidence

First agree the minimum acceptable source and annotation scope. Next inspect an appropriately accessible pilot selected to include difficult material. Review errors by category, decide which are repairable and record the accepted denominator. Only then extrapolate preparation work or negotiate a larger delivery.

Use milestone definitions in a proposal: source review, pilot, revised specification, batch preparation and release acceptance. State who resolves disagreements and how replacements are counted. If evaluation metrics are part of acceptance, define the evaluation set and scoring convention in advance; a data supplier should not promise an unmeasured model improvement.

Send Fiund the completed requirement fields and the unknowns you need resolved. We can review source fit and preparation scope for a project-specific proposal. For language coverage use the multilingual solution; for timing requirements read the full-duplex guide.

Questions buyers ask