# Speech data licensing

> Speech data licensing means giving an AI company a limited right to train on recordings you already hold – support and sales calls, podcasts, interviews or other voice archives. We advise you and negotiate the license, and you keep ownership and approve every deal.

By [Linus Näslund](https://corpusmarket.com/about.md), Co-Founder

## What speech data licensing means for the company that holds the recordings

Speech recognition, voice assistants and AI agents that answer the phone all learn from recorded speech. Most writing about speech data is aimed at the AI teams that need it. This page is for the other side: a company that already has recordings and wonders whether they can be licensed.

With a speech data license you don't hand over the archive for good. You grant an AI company a right to train on a defined set of recordings, on agreed terms, and you remain the owner. It is the same model as other [data licensing](https://corpusmarket.com/data-licensing.md), with a few questions that are specific to voice.

## Where AI companies get speech training data

AI companies get speech data from roughly four places:

- **Commissioned recordings** – paid speakers read scripts or hold prompted conversations, recorded to the AI company's specification.
- **Ready-made datasets** – data providers sell catalogs of speech that has already been collected and transcribed.
- **Open datasets** – research collections such as Mozilla's Common Voice, built from volunteer recordings.
- **Licensed archives** – real recordings held by companies, publishers and producers, licensed with documented rights.

A company with its own archive fits in the last group. Real calls and conversations show how people actually speak – interruptions, accents, background noise and the vocabulary of one industry – and that is hard to recreate with scripts.

## What speech data can be licensed?

- **Support and sales calls** – recorded phone conversations from a contact center or sales team, often with tickets and outcomes attached. See [customer service data](https://corpusmarket.com/sell-customer-service-data.md).
- **Meetings and internal calls** – recorded video meetings and conference calls, where participants were told about the recording.
- **Podcasts and radio** – episodes, raw interview tape and broadcast archives.
- **Interviews and research recordings** – market research, user interviews and oral history.
- **Lectures and webinars** – recorded teaching, talks and training sessions.
- **Voice-over and narration** – audiobooks, e-learning and commercial narration you commissioned.

Speech is one of several content types that can be licensed. Images, video and published text are covered on our page about [text, image and audio data](https://corpusmarket.com/sell-text-image-and-audio-data.md).

**[See what your data is worth](https://corpusmarket.com/submit-data)** – Free estimate in two minutes – not an offer

## What makes a speech archive easier to license

Buyers ask much the same questions about every speech archive:

- **Natural conversation** – real dialogue between people is generally in higher demand than text read aloud.
- **Hours and speakers** – volume is usually counted in hours of audio, and many different voices matter as much as many hours.
- **Language and domain** – what language is spoken and what field the conversations come from. What we see in the market right now is that demand is strongest for English.
- **Channels** – whether each side of a call is recorded on its own channel or mixed into one. Separate channels make it clear who says what.
- **Transcripts and labels** – existing transcripts, speaker turns, timestamps, language and date.
- **Original files** – the original recording format and quality, and whether anything has been edited or compressed.
- **Linked records** – tickets, outcomes or quality scores connected to the calls show what the conversation led to.
- **Consent records** – what each recording's notice or agreement says, not just that a call was recorded.

Audio, transcripts and linked records are assessed separately. A transcript with names removed does not mean the audio is de-identified – the same details are still spoken on the recording.

## Consent, voice and personal details

Speech raises questions that text does not, because a voice can identify a person. These are the ones that usually come up in the US:

- **Recording consent.** Federal law generally allows a call to be recorded when [one party consents](https://www.law.cornell.edu/uscode/text/18/2511). A group of states, California among them, [require every party's consent](https://www.dmlp.org/legal-guide/recording-phone-calls-and-conversations), and calls often cross state lines. So what callers were told, and where they were, is part of every review.
- **Purpose of the notice.** A notice that a call is recorded for quality or training purposes was written for internal use. Whether it covers licensing to an AI company is looked at case by case. The FTC has said that [quietly changing terms](https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/02/ai-other-companies-quietly-changing-your-terms-service-could-be-unfair-or-deceptive) to allow new uses of data, such as AI training, could be unfair or deceptive.
- **Voiceprints.** Illinois' [Biometric Information Privacy Act](https://www.ilga.gov/legislation/ilcs/ilcs3.asp?ActID=3004&ChapterID=57) lists a voiceprint among its biometric identifiers, requires written notice and a written release before one is collected, and restricts selling or profiting from one. Whether a recording is used in a way that involves voiceprints depends on what the buyer does with it, which is one reason licenses spell out permitted uses.
- **Voice as likeness.** Some state laws protect a person's voice alongside their name and likeness – California's Civil Code section 3344 is one example. This matters most for hosts, narrators and voice actors, and for any use that would imitate a recognizable voice.
- **Personal details in the audio.** Names, addresses and account details are spoken aloud in many calls. Card numbers and other sensitive details are filtered out before delivery, and the plan for [de-identification](https://corpusmarket.com/guides/anonymized-vs-de-identified-data.md) covers both audio and transcripts.
- **Whose recordings they are.** If calls were handled on behalf of a client, or a podcast was produced with a network or distributor, the contract may decide who can license the recordings.

This describes how things generally work today and is not legal advice. Every archive is different, so we look at each case on its own and bring in specialists when a question calls for it. There is more background on our page about [data privacy and AI training](https://corpusmarket.com/data-privacy-and-ai-training.md).

## How a speech data license is set up

By default the license is **non-exclusive** and limited in scope, time and use. For speech, the use is often the most important part:

- **Permitted uses** – for example training speech recognition or a voice agent, as opposed to generating a synthetic voice that sounds like a particular speaker.
- **Scope** – which recordings, which years, and whether transcripts and linked records are included.
- **Term** – how long the buyer may use the recordings.
- **No resale** – the buyer may not pass the recordings on.

Because the license is non-exclusive, the same archive can in some cases be licensed to more than one buyer. What a speech archive may be worth depends on the material and on demand at the time – see [what your data is worth](https://corpusmarket.com/what-is-your-data-worth.md). We never promise a price before a buyer has made a written offer.

## How it works

1. **You describe the recordings** in our [form](https://corpusmarket.com/submit-data.md) – type of speech, approximate hours, languages and where the recordings come from. No files needed.
2. **We assess** rights, consent, personal details and demand, and get back to you within 48 hours on business days.
3. **You approve** a concrete offer and agreement – or say no. Nothing is shared until you approve the next step.
4. **Delivery and payment** take place according to the agreement, after any de-identification.

Buyers see an anonymous profile first, and you approve every step after that. If recordings are one part of a larger picture – email, chat, CRM and files from the same team – see how to [sell data to AI companies](https://corpusmarket.com/sell-data-to-ai-companies.md).

## Frequently asked questions

### What is speech data licensing?

It is an agreement that lets an AI company train on recorded speech that someone else holds – calls, interviews, podcasts or other voice recordings. The holder keeps ownership and grants a right of use. By default that right is non-exclusive and limited in scope, time and use.

### Can we sell call recordings to AI companies?

Sometimes. In practice it is a license rather than a sale, and what usually decides it is who controls the recordings, what callers were told when the call was recorded, which states the callers were in and how personal details are removed from the audio and the transcripts. Each archive is reviewed on its own.

### Where do AI companies get speech training data?

Broadly from four places – recordings commissioned from paid speakers, ready-made datasets from data providers, open research datasets, and licenses for real recordings that companies and publishers already hold. A company with its own call or audio archive belongs to the last group.

### What speech data do AI companies look for?

Mostly natural conversation between real people, with many different speakers and the vocabulary of a specific field. Transcripts, speaker labels, separate channels for each side of a call and clear documentation of consent make an archive easier to use.

### Do the speakers have to consent before recordings are licensed for AI training?

It depends on the recordings, the states involved and what the speakers were told. A notice that a call is recorded for quality purposes does not necessarily cover a new use such as AI training, so the actual notices and agreements are reviewed case by case before anything is offered.

### Is licensed speech data cleared for commercial use?

That is set by the license, not by the recordings. An agreement states which uses are allowed – for example training speech recognition for a commercial product – and which are not, such as creating a synthetic copy of a speaker's voice. The documented rights behind the recordings are what the buyer relies on.

### Can a podcast or interview archive be licensed for AI training?

Often, when you hold the rights to the recordings. Guests, co-hosts, freelancers and licensed music can each come with their own agreements, so those are checked first and only what you have the right to license is included.

### Does it cost anything to have a speech archive assessed?

No. The assessment is free and you don't send any files. We only get paid when a deal goes through, through a share of the revenue.