# How to license video and audio archives for AI training

URL: https://sellmydatatoai.net/guides/sell-video-and-audio-for-ai-training
Reviewed: 2026-11-08

You can license video and audio archives for AI training if you hold every rights layer in the recordings: the master, the music, the performances and the people on screen or on mic. Bloomberg reported in January 2025 that AI companies were paying $1 to $4 per minute of unused video. Clear rights, original masters and searchable metadata decide what an archive earns.

## Key takeaways

- Reported video rates run from about $1 to about $7.25 per minute, according to Bloomberg's reporting on OpenAI, Google, Moonvalley and Adobe.
- Owning the master is not enough. Music, performers, people in shot and third-party clips are separate rights layers.
- Voice and faces can be biometric data, and Illinois' BIPA bars selling or otherwise profiting from voiceprints and scans of face geometry.
- Start with rushes, b-roll and in-house recordings you own outright, then clear or remove the rest.
- Deliver original masters with per-clip metadata and transcripts. Padded or compressed minutes are worth less.

Video and speech models need very large volumes of recordings, and the developers building them now pay for licensed footage and audio. Bloomberg reported on January 10, 2025 that OpenAI, Google and Moonvalley were among the AI companies paying creators for unused video, at $1 to $4 per minute of footage.

The market is not new. Bloomberg Law reported that Shutterstock's AI licensing business generated $104 million in 2023, and Shutterstock signed a six-year agreement with OpenAI covering images, videos and music in July 2023, according to TechCrunch. This guide covers who buys recordings, what they pay, which rights you need and how to prepare an archive. It describes what the law says, not how it applies to your archive, so take independent legal advice before you license.

## Who buys video and audio for AI training?

Video and audio for AI training is bought by model developers directly and through licensing agents, platforms, brokers and marketplaces that assemble datasets for them. The table shows the main routes, using what each company says on its own site.

| Route | Example | What it says it handles |
| --- | --- | --- |
| Licensing agent | Troveo | Video, audio and gaming datasets, citing more than 8 million hours of video and 4 million hours of audio |
| Video licensing platform | Versos | Indexes libraries in place in cloud storage such as S3 or Cloudflare and delivers datasets to the licensee |
| Data broker | Protege | Video, audio and speech, healthcare and other domain data for model builders |
| Data marketplace | Defined.ai | Speech, text, image, video, music and sound effects datasets from vetted partners |
| Training-data company | Labelbox Prism | Consented likeness and voice data, among other categories |
| Content licensing platform | Shutterstock | Its own contributor library of images, video, music and 3D |

Troveo's homepage says it sources from a network of more than 7,000 data providers and that 95% of its content is exclusive to Troveo. Versos's list of places to sell video for AI training names Google, Meta, NVIDIA, Runway, Moonvalley, Synthesia and Getty Images. Our profiles of [Troveo](https://sellmydatatoai.net/data-buyers/troveo) and [Versos](https://sellmydatatoai.net/data-buyers/versos) set out the published terms of each.

## How much do AI companies pay for video and audio?

AI companies pay for video by the minute, and the public figures run from about $1 to about $7.25 per minute. Published prices for audio are much rarer.

| Buyer | Reported rate | Source |
| --- | --- | --- |
| OpenAI, Google, Moonvalley and others | $1 to $4 per minute of unused footage | Bloomberg, January 2025 |
| Adobe | Average about $2.62 per minute, up to about $7.25 | Bloomberg, April 2024 |
| OpenAI | $1 to $4 per minute | Versos blog |
| Adobe | About $3 per minute | Versos blog |

The arithmetic shows why rights matter at scale. An archive of 10,000 hours holds 600,000 minutes, so the reported $1 to $4 range implies $600,000 to $2.4 million before any footage is excluded. That is a calculation on reported rates, not a quote, and buyers discount footage with open rights questions.

For company audio, the closest published figure is Mercor's listing for call centers: $20,000 to $250,000 per company for operations with 50 to 5,000 agents, as recorded by ddiy.co in October 2026. Every published figure on our site is collected in the [AI data price index](https://sellmydatatoai.net/ai-data-price-index).

## What rights do you need to license recordings for AI training?

To license recordings for AI training you need rights to every layer in them, not just ownership of the master file. US copyright law treats the layers as separate works: section 102(a) of the Copyright Act lists musical works, dramatic works, motion pictures and other audiovisual works, and sound recordings as separate categories.

| Layer | Usually owned by | What to check |
| --- | --- | --- |
| Footage or audio master | The company that shot or commissioned it | Commission, production and terms of trade contracts |
| Music | Publishers for the composition, labels for the recording | Whether the license covered only the original production |
| Performances | Performers | Whether their consent reaches new uses |
| People on screen or on mic | The individuals | Releases, privacy notices and biometric rules |
| Inserted clips and stock | The agency or library | The agency's license terms |
| Logos and artworks in shot | Third parties | Whether to blur, crop or exclude |

Getphyllo's guidance on training data licenses calls this per-layer clearance, because video contains music, images, performances and trademarks that can each belong to someone different. Publicity laws add a layer for the people on screen and on mic. California Civil Code section 3344 makes anyone who knowingly uses another's name, voice, signature, photograph or likeness on products or in advertising without consent liable for the greater of $750 or actual damages, which matters if a buyer may generate outputs that imitate them.

Raw rushes are often simpler to license than finished programs, because the music, graphics and archive inserts are added in the edit. Our page for [media companies and video archives](https://sellmydatatoai.net/sell-your-data/media-and-video-archives) covers commissioned programs and agency footage in more detail.

## Do speakers and people on screen need to consent?

Whether speakers and people on screen need to consent depends on where they were recorded, the lawful basis relied on and what the buyer will train. The rules tighten sharply when voices or faces are processed to identify people.

- Illinois' Biometric Information Privacy Act lists a voiceprint and a scan of face geometry among its biometric identifiers, requires a written release before they are collected and bars selling, leasing, trading or otherwise profiting from them, with liquidated damages of $1,000 per negligent violation and $5,000 per intentional or reckless one.
- Federal wiretap law allows a recording where one party to the conversation consents, under 18 U.S.C. 2511(2)(d), but some states set a stricter rule.
- California Penal Code section 632 makes it an offense to record a confidential communication without the consent of all parties, so a recording must have been lawfully made before it can be lawfully licensed.
- Tennessee's ELVIS Act, signed on March 21, 2024 and in force from July 1, 2024, added voice to the state's publicity rights, covering both a real voice and a simulation of it, according to Manatt.

Buyers reflect these rules. Labelbox Prism says likeness and voice data need explicit, timestamped consent. If a buyer plans voice cloning or speaker identification, agree those uses separately in the license, or exclude them. Our page on [speech and audio data](https://sellmydatatoai.net/buy-ai-training-data/speech-and-audio-data) lists what buyers check.

## Can you license call recordings and meeting audio?

You can license call recordings and meeting audio only where they were lawfully recorded and the people on them were told about uses that cover AI training, or the personal data is removed. For most company audio that means licensing redacted transcripts and dropping the audio itself.

Demand exists. Polyshares lists call recordings among the operational records it licenses, and Mercor publishes a range for call centers. The risk is concentrated in the content: customers read out card numbers, addresses and health details. Cut payment segments, redact spoken identifiers in the transcript and treat the voice itself as personal data unless you can show otherwise.

## How do you prepare a video or audio archive for buyers?

You prepare a video or audio archive by sorting it by rights position first, then describing what remains so a buyer can search and price it. The order below avoids clearing material you will later exclude.

1. **Inventory the archive**: Record total hours by format, year, resolution, language and storage location, and separate original masters from compressed copies.
2. **Sort by rights position**: Group material into owned outright (rushes, b-roll, in-house recordings), cleared with conditions, and uncleared. Offer the first group first.
3. **Remove or flag third-party layers**: Strip music beds, agency inserts and stock clips you do not control, or deliver muted or stem-separated audio where the buyer accepts it.
4. **Deal with the people**: Match releases and consents to clips, and exclude footage of identifiable people where you hold neither a release nor a lawful basis that covers training.
5. **Add searchable metadata**: Write shot descriptions, time-aligned transcripts and speaker labels. Versos notes that without a searchable structure even valuable assets can stay hidden.
6. **Count minutes honestly**: Remove duplicates, slates, black frames and long silences before counting, because buyers price per useful minute.
7. **Build the manifest and sample**: List hours, formats, rights per group and exclusions, then draw a random sample. Our guide to [building a data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest) sets out the fields.

## What formats do AI buyers expect for video and audio?

AI buyers expect video and audio as original-resolution or lossless masters with per-clip metadata, not compressed web copies. The table reflects what buyers ask for on our [video training data](https://sellmydatatoai.net/buy-ai-training-data/video-training-data) and speech pages.

| Asset | Preferred delivery | Metadata buyers ask for |
| --- | --- | --- |
| Video | MP4, MOV or ProRes masters at original resolution | Duration, resolution, frame rate, codec, shot description |
| Audio | WAV or FLAC lossless files | Sample rate, channels, speaker IDs, language, recording environment |
| Transcripts | Time-aligned JSON, SRT or TextGrid | Speaker labels and accuracy measured on a sample |
| Rights records | One record per asset | Releases, consents, licenses in and exclusions |

## Should you license footage exclusively or through an agent?

Licensing footage exclusively pays more per deal, while an agent or a non-exclusive license lets the same minutes earn more than once. Troveo pays a revenue share on every sale with no fees and asks for time-limited exclusivity, according to its business data page. A direct license with a model developer is usually a per-minute or fixed fee negotiated deal by deal.

Check what you have already granted. YouTube's help page says channel owners and rights holders can allow third-party companies to train AI models on their public videos, and that the setting is off by default. If you turned it on, a buyer asking for exclusivity needs to know. Our guide to [exclusive and non-exclusive licenses](https://sellmydatatoai.net/guides/exclusive-vs-non-exclusive-data-licenses) sets out the trade-off.

## If you hold UK or EU data

Recordings made in the UK or EU, or featuring people there, carry extra rights layers and stricter rules on voices and faces.

- Separate works: section 1(1) of the Copyright, Designs and Patents Act 1988 lists musical and dramatic works separately from sound recordings, films and broadcasts.
- Performers: [section 182 of the same Act](https://www.legislation.gov.uk/ukpga/1988/48/section/182) makes recording a live performance without the performer's consent an infringement, and the consent in a performer's contract may be limited to broadcast or distribution.
- Biometric data: UK GDPR Article 4 defines biometric data as personal data from specific technical processing that allows or confirms the unique identification of a person, and Article 9 makes biometric data processed for that purpose special category data.

## What should a video or audio license say?

A video or audio license should say which model uses are allowed, which are excluded and what happens if a rights holder later objects. Daniel Goldberg of Frankfurt Kurnit Klein & Selz notes that once data is used to train a model, "it cannot practically be pulled back out", so these terms have to be settled before transfer.

- Permitted uses, such as video generation, recognition or captioning
- Voice cloning, speaker identification and likeness generation excluded unless separately consented
- Outputs that reproduce an identifiable performer or contributor prohibited
- Deletion of raw copies by a fixed date
- Withdrawal terms covering future training runs and models already trained, which Getphyllo lists among six things a training data license should cover

We do not take custody of your recordings. Buyers see redacted samples under a mutual NDA, and full files move only after you sign a license directly with the buyer. If you hold an archive and want to know which parts buyers would take, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller).

## Frequently asked questions

### How much is an hour of video worth to AI companies?

At the $1 to $4 per minute Bloomberg reported in January 2025, an hour of licensable footage works out at $60 to $240. Bloomberg reported in April 2024 that Adobe paid an average of about $2.62 per minute, rising to about $7.25. Footage with uncleared music, people or inserts earns less or is excluded, so the figure applies to cleared minutes only.

### Can I license footage I have already posted on YouTube?

Possibly, if you own every rights layer in it. YouTube's help page says creators and rights holders choose whether third-party companies may train AI models on their public videos, and the setting is off by default. A direct license to a buyer is a separate deal. Tell the buyer about any training permission you have already given, especially if they want exclusivity.

### Does music in my footage need separate clearance?

Usually yes. Music has its own copyrights in the composition and the sound recording, and sync licenses are normally limited to the original production. Getphyllo's guidance calls for per-layer clearance for this reason. Many sellers deliver muted or stem-separated audio, or offer rushes recorded before music was added in the edit.

### Is recorded speech biometric data?

Recorded speech is personal information where a speaker is identifiable, and it can become biometric data once processed to identify or verify someone. Illinois' Biometric Information Privacy Act lists a voiceprint as a biometric identifier and bars profiting from it, and UK GDPR Article 9 treats biometric data used for identification as special category data. Take advice before licensing audio for voice cloning or speaker identification.

### Can a business license its customer call recordings?

Only with care. The calls must have been lawfully recorded, and the privacy information customers received must cover the new use, or the personal data must be removed. Most sellers license redacted transcripts rather than audio, cut payment segments and exclude calls with health or financial detail. California Penal Code section 632 shows how strict recording consent rules can be.

## Sources

- [YouTubers Are Selling Their Unused Video Footage to AI Companies](https://news.bloomberglaw.com/ip-law/youtubers-are-selling-their-unused-video-footage-to-ai-companies), Bloomberg Law, accessed November 2026
- [Adobe is buying videos for $3 per minute to build AI model](https://www.bnnbloomberg.ca/adobe-is-buying-videos-for-3-per-minute-to-build-ai-model-1.2057606.amp.html), Bloomberg via BNN Bloomberg, accessed November 2026
- [Shutterstock's AI-licensing business generated $104 million](https://news.bloomberglaw.com/ip-law/shutterstocks-ai-licensing-business-generated-104-million), Bloomberg Law, accessed November 2026
- [Shutterstock expands deal with OpenAI to build generative AI tools](https://techcrunch.com/2023/07/11/shutterstock-expands-deal-with-openai-to-build-generative-ai-tools), TechCrunch, accessed November 2026
- [Top 10 places to sell your video content for AI training data](https://www.versos.ai/blog/sell-video-content-ai-training-data), Versos, accessed November 2026
- [Troveo | Real-World Data Powering Top Frontier Labs](https://www.troveo.ai/), Troveo, accessed November 2026
- [Business data](https://www.troveo.ai/business-data), Troveo, accessed November 2026
- [Prism](https://prism.labelbox.com/), Labelbox, accessed November 2026
- [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026
- [17 U.S. Code section 102: subject matter of copyright](https://www.law.cornell.edu/uscode/text/17/102), Legal Information Institute, Cornell Law School, accessed November 2026
- [California Civil Code section 3344](https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=CIV&sectionNum=3344), California Legislative Information, accessed November 2026
- [Biometric Information Privacy Act, 740 ILCS 14](https://www.ilga.gov/legislation/ilcs/ilcs3.asp?ActID=3004&ChapterID=57), Illinois General Assembly, accessed November 2026
- [18 U.S. Code section 2511: interception and disclosure of wire, oral, or electronic communications prohibited](https://www.law.cornell.edu/uscode/text/18/2511), Legal Information Institute, Cornell Law School, accessed November 2026
- [Copyright, Designs and Patents Act 1988, section 1](https://www.legislation.gov.uk/ukpga/1988/48/section/1), legislation.gov.uk, accessed November 2026
- [Copyright, Designs and Patents Act 1988, section 182](https://www.legislation.gov.uk/ukpga/1988/48/section/182), legislation.gov.uk, accessed November 2026
- [Best public data providers for AI training and grounding](https://www.getphyllo.com/post/best-public-data-providers-ai-training-grounding), Getphyllo, accessed November 2026
- [UK GDPR Article 4: definitions](https://www.legislation.gov.uk/eur/2016/679/article/4), legislation.gov.uk, accessed November 2026
- [Biometric Information Privacy Act, 740 ILCS 14/10 definitions](https://lig.ilga.gov/legislation/ilcs/documents/074000140K10.htm), Illinois General Assembly, accessed November 2026
- [California Penal Code section 632](https://law.justia.com/codes/california/code-pen/part-1/title-15/chapter-1-5/section-632), Justia, accessed November 2026
- [Tennessee's ELVIS Act Expands Publicity Rights for Individuals' Voices & Protects Against AI Misuses](https://manatt.com/insights/newsletters/client-alert/tennessees-elvis-act-expands-publicity-rights-for), Manatt, accessed November 2026
- [Your content & third-party training](https://support.google.com/youtube/answer/15509945), YouTube Help, accessed November 2026
- [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein & Selz, accessed November 2026

---

Book a free consultation: https://sellmydatatoai.net/book-a-consultation