Why AI buyers want it
Labs want their voice products to work for speakers they do not yet serve. Natural, unscripted speech from many speakers, with transcripts, is what closes the gap. Organizations with consented speech in regional languages, dialects and accents hold some of the scarcest audio in AI.
Speech data is like a library: the shelves for major languages are full, and the shelves for rare ones are nearly empty.
What sets the price
- Language rarity.
- Natural, unscripted speech.
- Many speakers across ages and regions.
- Accurate transcripts.
- Consent that names AI training, tied to each file.
The estimate on this page prices consented, transcribed speech across many low-resource languages. Raw audio is worth less until it is transcribed.
What blocks a deal, and the fix
- Missing speaker consent
- Consent forms matched to each file; we help collect them from speakers you can still reach.
- Voiceprint laws
- Biometric privacy rules apply in some regions, so we screen each set against them before scoping.
What owners overlook
- Rare languages are listed at $60 to $150 per hour, against $50 to $100 for ready-made sets in major languages.
- Tutoring sessions and language lessons are strong speech data when consent covers them.
- Metadata on each speaker's age, region and accent can matter as much as the audio.
Market evidence
- A speech data buying guide lists ready-made speech in major languages at $50 to $100 per hour and prices custom collection higher still.SpeechData.ai, (opens in a new tab)
- Open speech recognition models now reach far beyond the major languages, into languages with little recorded speech.VentureBeat, (opens in a new tab)
How a deal works
Appraise
Answer a few questions about your recordings. See what affects the price and whether an indicative range is available. No files, no access.About a minute
Verify title
We confirm who holds the rights to each recording and whether consent covers AI training, and plan the masking of names and other identifiers.Before any offer
Route
We match the data to the license routes that fit it and show you the terms of each before you decide. Nothing is offered without your approval.Sign and get paid
You approve the agreement and its scope before anything moves. We never hold your passwords. Transfers use scoped, revocable connections your admins approve, or exports.Confirmed after review
Questions
What is speech data worth per hour?
A published price guide lists ready-made speech datasets at $50 to $100 per hour in major languages and $60 to $150 in low-resource languages. For 2,000 hours at $60 to $150 per hour, our estimate is $120,000 to $300,000 at Medium confidence. Indicative estimate, not an offer.
How many hours do we need?
There is no published minimum. Tell us your hours in the appraisal and we will say whether a route fits.
What consent do we need?
Consent from each speaker that names AI training, tied to each file. Consent to record or to teach is not enough for licensing, and biometric laws apply in some regions.
Do we need transcripts?
They raise the price. Accurate transcripts, speaker labels and metadata on age, region and accent turn raw audio into a dataset buyers can use straight away. Ready-made prices assume transcripts are included.
Can we license the same recordings more than once?
It depends on the route. Ready-made datasets are often licensed to several buyers, while a custom collection contract may set exclusivity for a term. The terms are agreed in writing, and you approve the agreement and its scope first.
What do we warrant when we license speech?
That every speaker gave consent that covers AI training and that you hold the rights to the recordings. Only include files whose consent you can show.
