Sell Speech Data

What buyers ask a speech supplier for

Speech data demand is precise, and the specification almost always contains the same fields: language and dialect, total hours, minimum distinct speaker count, recording conditions, annotation depth, and delivery format. Speaker count is the field suppliers underestimate. Hours are easy to add by recording the same voices longer; a buyer who asks for 100 hours across 80 speakers is asking for something harder, and that difficulty is where your margin lives.

The second most common requirement is conditions. Studio recording with a low noise floor, home recording, mobile capture, or a defined noisy environment — each is a different product, and a supplier who can only deliver one of them is limited to the projects that match it.

One structural fact shapes everything below: there is no public portal at any major AI company where speech data can be uploaded and paid for. OpenAI, Google, and Anthropic buy through data teams and vendors. Demand reaches a supplier as a specification, which is why the rest of this page is about specifications and rights rather than about where to upload. And the specification is not a formality: audio that matches no live requirement cannot be paid for, no matter how clean the recording is.

What you need before you can sell speech

Four capabilities, and buyers check all of them.

• Native-speaker access — you can recruit people who genuinely speak the target language or dialect, in the numbers the specification needs.
• A recording setup that meets the condition being sold — microphone, room treatment, and a fixed signal chain, applied consistently across all sessions.
• Annotation capability, or a partner who has it — transcription at minimum, and ideally the deeper layers: intent, diarization, phoneme alignment.
• A consent process — a signed release from every speaker, specific enough about AI training use that a buyer's legal review will accept it.

The consent question, in voice

Voice data is personal data, and a recording is the person in a way that a text file is not. Every speaker in your dataset needs a document that says what the recording will be used for, who it may be shared with, how long it is retained, and how they can withdraw. This is the most common reason speech suppliers are turned away — not audio quality, not price, but a rights chain that cannot be shown.

It also defines what you can sell. You can sell recordings you produced, of speakers who consented. You cannot sell recordings of people who did not, and there is no intermediary who can fix that after the fact.

What is not sellable, and how the money works

Three categories are declined by us and by serious buyers everywhere. Recorded telephone calls, because the other party on the call never consented. Medical and clinical recordings, because consent and regulation make them a different business entirely. And audio scraped from podcasts, videos, or social media that you do not own — public availability is not a license, and rights holders can and do pursue it.

On payment: money follows acceptance. You deliver, the buyer's QA reviews against the criteria written at the start, and the batch is paid when it passes. A supplier who prices to survive one round of re-work is safe; one who treats handover as the finish line is not.

If your material is in the sellable categories and you can document your rights, the practical step is a specific first message: which languages and dialects you can produce, your recording conditions, your annotation depth, and your monthly capacity in hours. That is what can be matched against a live requirement — and if there is no match today, it tells us what to bring you next. Send it through the request form, selecting custom data collection, and it goes on file against the requirements we are working.