Selling audio and speech recordings for AI training

AI companies buy speech by the hour, and what an hour fetches depends on which side of the market it sits: speech recorded to a buyer's order is the dearest audio there is, and speech that already sits in a company's archive is the cheapest. Across every public price on record, licensed speech sold for a median of $76 per hour of audio, ready-made speech datasets ask a median of $118 per hour of audio, and speech recorded to a buyer's specification costs a median of $122 per hour of audio, while the people who record speech for collectors are paid a median of $24 per hour of audio. The one buyer that publishes a price for audio a holder already has, Shaip, offers up to Rs. 200 per audio hour, about $2.11, for existing podcasts, radio programmes and news broadcasts. Whether a recording can be sold at all is a question of consent: federal law lets one party record a call, six states require every party for calls and conversations alike, and no statute we read says whether a lawfully made recording needs fresh consent before it trains a model.

This guide is for any company that holds recorded speech: a contact centre or a sales team with years of calls, a podcast network or a broadcaster with an archive, a firm whose meetings and voice notes were recorded, a research group or language community with field recordings. It is also for people who record speech for pay, because what they are paid is one end of the same market. It covers who buys audio and what for, what each kind of audio is worth on each side of the market, why archive audio sells cheapest, the recording-consent law of the fourteen states whose statutes, read for this guide, depart from the federal rule, the open question of consent for AI training, de-identifying voices and transcripts, what buyers' terms ask, how to prepare an archive, and how a sale through AIData.Broker works. Every number links to its record, and every statute to its text.

AIData.Broker keeps the price board, the rates page and the audio and speech page this guide draws on, and it also acts as the agent of holders who license or sell their data, so we are a party in this market. Check any figure you plan to rely on against its source.

This guide is research from public sources and the statutes themselves, not legal advice.

Contents

1. Who buys audio, and what for
2. What each kind of audio is worth, on each side of the market
3. Why archive audio is the cheapest audio
4. Recording consent, state by state
5. The gap: consent to record is not consent to train
6. De-identifying voices and transcripts
7. What buyers' terms ask of audio
8. How to prepare an archive for sale
9. How a sale through us works
10. Questions people ask
11. What this means for you

Who buys audio, and what for

Speech is bought to teach machines four different things, and each wants different audio. Knowing which one a buyer is building tells a holder which part of an archive it will pay for and which it will pass over.

Speech recognition (turning speech into text) needs volume and variety: many speakers, accents, microphones, rooms and topics, with accurate transcripts. It is the oldest market and the largest by hours. The clearest record of it is the audited prospectus of Beijing Haitian Ruisheng Science Technology (SpeechOcean) (SpeechOcean), a Chinese speech data company, which reported that in 2020 it licensed 199,980 hours of its own off-the-shelf speech datasets, almost all of it recognition data, at an average of CNY 413 an hour.

Speech synthesis (text to speech, and the voices of assistants) needs the opposite: few speakers, clean studio sound, expressive and consistent delivery. The same prospectus shows synthesis data selling in far smaller volume at a far higher price, an average of CNY 13,832 an hour. Voice companies still buy it: Murf says it is "procuring voice data in multiple languages and accents to train our base models", with its focus on "conversational-style data" for podcasts, audiobook narration and voiceovers, and a project on Twine paid voice actors $100 per finished hour of audio for training data. A separate market pays people for the use of their voice itself: on ElevenLabs' voice library the person whose voice is cloned earns $0.03 per 1,000 characters generated by default, and voice creators had earned over $22 million there by May 2026. That is a royalty for a voice, not a fee for training data.

Voice agents (systems that hold a spoken conversation, take a call or book an appointment) need real two-way conversation: people talking over each other, pausing, correcting themselves and responding naturally. Ocular AI, a voice-data start-up reported to sell to frontier labs, is described as supplying " material covering pauses, interruptions, overlapping speech and natural response". This is why call-centre and sales-call archives are now in demand: Fiund, a company that arranges licences of call recordings and other archives, writes that task-oriented, multi-speaker phone audio is "valuable for voice agents and ASR", and in one reported purchase a lab buyer took a company's corpus that included about 83K sales-call recordings beside its code, tickets and documents. Collectors pay more for conversation than for reading: $28 an hour for an unscripted two-person conversation at Perit AI, against a base of $20 per approved audio hour for read speech.

Language coverage is the fourth. Models are weakest in the languages and accents the large datasets miss, and buyers pay to close the gap. The co-founder of Karya, which supplies speech data to large technology companies, said an hour of Odia speech " used to cost about $3-$4, now it's $40". Protege's speaker tracks cover more than 70 languages, Sieve lists multilingual recordings beside podcasts and audiobooks among the speech it would license, and while Shaip's own catalogue already holds 217,795 hours of English speech, its published offer for existing audio is for Hindi, Marathi, Telugu and other Indic languages.

Who buys, and how they pay

Most holders will not sell to a model developer directly. As with video, the labs prefer large, cleared, described volumes, so speech reaches them through data companies that collect it, buy it, license it on, or broker it. The table lists the buyers and intermediaries whose own pages say how they take speech; the buyer board keeps their current terms.

Two things in the table matter for a holder. Only one of these buyers publishes a price for existing audio, so in every other case the number comes from a conversation, which is why the market figures below matter. And the terms differ more than the headlines: a training-only licence that leaves ownership with the holder (Shaip, Protege) is a different thing from a perpetual, irrevocable, sublicensable one (Kled), which is close to selling the training rights outright.

What each kind of audio is worth, on each side of the market

Audio is priced per hour of recorded speech. The same hour carries very different prices depending on where in the market it is quoted: what a person is paid to record it, what a buyer offers a holder for it, what a dataset asks off the shelf, what a buyer pays a vendor to record it to order, and what licences that actually happened came to. Reading a price without knowing its side is the most common mistake in this market. The rates page computes each side from every public price at once.

By kind of audio

Within each side, the kind of audio moves the price more than anything else. The table gives the public figures for the kinds of speech companies actually hold, beside what speech recorded to order costs.

The gaps in that table are not noise. An hour of transcribed, two-channel conversation with documented consent asks many times what an hour of bulk call audio asks, and the difference is almost entirely the work and the rights, not the speech. A vendor's own comparison puts off-the-shelf call-centre audio at $5 to $25 per hour against $50 to $300 per delivered hour for custom collection, and the top of the custom range for speech runs 2.7x the top of the catalogue range.

The chart puts the main figures on one axis, per hour of audio. The point it makes is that the further audio sits from the speaker's consent and the buyer's specification, the less it fetches.

Sources: existing Indic audio, bulk podcast audio, call-centre audio from corpora, low-resource conversation, speech recorded on phones, real support calls and their usage rate, conversational datasets, speech to order, studio voice data.

If you record speech for pay

People who record speech for collectors are paid per approved hour of audio, per hour of work, or per short task, and the published rates are close together. Perit AI pays a base of $20 per approved audio hour, more for prompted speech, noisy rooms and conversation, and up to $32 for accents it is short of. Luel pays $12 an hour for American English conversation and $7.50 an hour for conversations in low-resource languages such as Uyghur and Tibetan. Shaip pays Haryanvi speakers up to Rs. 500 an hour, Dhivehi speakers $20 for transcribing an hour of audio and $15 for reading 600 sentences, and pays per session elsewhere ( $10 for about 45 minutes of Sylheti, $8 for 16 phrases in European languages, Rs. 500 for a wake-word session). Jobs paid by the hour of work list $10 to $50 (Pila8) and $15 to $40 (YPAI). Across all of them the median is $24 per hour of audio per hour of audio and $24 per hour of work per hour of work.

Professional voice work pays far more, and its terms matter as much as its rate. The Twine project paid $100 per finished hour and promised the voices would not be used to create an AI voice likeness; a Voices.com job pays $120 to $400 per hour of finished audio, the highest rate in this research, but for a dataset licensed for evaluation only, not for training or cloning. Evaluation licences are the cheaper kind on the buyers' side: in the data catalogues we read that list both, a narrower licence such as evaluation is priced at about 0.2 to 0.5 times full commercial use, so a high rate for evaluation audio is not a benchmark for training audio.

Why archive audio is the cheapest audio

The one published offer for audio a holder already has, Shaip's up to Rs. 200 per audio hour, sits below every rate paid to people who record speech to order, from Luel's $7.50 for low-resource conversation to Perit's $20 base, and far below catalogue datasets of real calls listed at $25 an hour. Bulk archives listed for sale say the same thing from the other side: podcast audio at about $4.86 an hour, Swahili call recordings at about $1.50, and one multilingual set of 35,000 hours from $20,000. It looks backwards that an hour someone has already recorded should be worth less than an hour someone still has to record. Five reasons explain it.

The consent was given for something else. A speaker recorded to order agrees, on the recording, to the use: Perit says consent " is captured on the recording, in plain language, and it names what the audio trains". A call recorded under a "this call may be recorded" notice, a podcast guest who agreed to an episode, or a meeting recorded for the minutes did not agree to that. Whether that matters in law is the open question in the next two sections; that it matters to a buyer's price is not, because the buyer either discounts it or declines.

It was not recorded for a model. Archives are mostly single-channel (both voices mixed on one track), compressed for storage, untranscribed and undescribed. Fiund's advice to holders is to export "original audio, dual-channel where it exists, not compressed archival copies", and to state hours, call counts, languages and the recording setup; the two catalogue datasets of real support calls above sell at their price because they are speaker-separated, redacted and transcribed.

General speech is plentiful. Listen Notes counts at least 3,824,033 podcasts, Spotify hosts 7 million podcast titles, and the volunteer Common Voice collection holds 42,593 hours in 295 languages. A buyer choosing between podcast archives pays the general rate and keeps its premium for what is scarce.

It carries personal information. Calls are full of names, account numbers, addresses and card details, spoken aloud. Removing them costs money before a buyer can use the audio, and some buyers refuse it outright: Murf asks holders " Please do not provide any personal data such as call recordings or any audio of non-consenting speakers".

Nobody publishes a rate for it. Fiund says " There is no standard rate card", and Murf, Protege and Troveo publish no rate for existing audio either. A holder with no reference number accepts the first one offered, and the first one offered is the buyer's.

None of this makes an archive worthless. It means its price is set by what can be done to it before sale. Each of the five reasons has a remedy (fresh consent or a lot limited to consented speakers, export of original two-channel audio, transcripts and speaker labels, removal of personal details, and competing offers), and the remedies are what move an archive from the bottom of the table toward the conversational datasets. An archive also has an advantage recording to order lacks: it can be licensed again. Haitian Ruisheng sold 576 licences on 395 of its own datasets in a single year, while data it recorded to a client's order passed to that client and could not be resold.

Recording consent, state by state

Before anything about AI, a recording has to have been lawfully made, because a defect in how it was made follows it into every sale. The federal Wiretap Act is a one-party rule: it is not unlawful for a person " to intercept a wire, oral, or electronic communication where such person is a party to the communication or where one of the parties to the communication has given prior consent", unless the purpose is criminal or tortious. A business recording its own calls is a party, so federal law is satisfied by its own consent. But the same Act separately punishes anyone who intentionally discloses or uses the contents of a communication knowing it was intercepted unlawfully, and states may be stricter than the federal floor. Several are.

The table covers the states whose statutes depart from the one-party rule, read in their own text. "All parties" means every person on the call or in the conversation must consent; "one party" means a participant may record. States not in the table were not read for this guide.

A few rows need more than a cell.

California has two rules, and the stricter one has no confidentiality condition. Its general section reaches only a confidential communication, which excludes any conversation where the parties may reasonably expect to be overheard or recorded. But its section on mobile and cordless phones forbids recording, " without the consent of all of the parties to a communication", any call with a cellular or cordless phone on one end, with no such exception. Since most customer calls touch a mobile phone, all-party consent is the basis a California call archive needs.

Notice works differently from state to state. In Washington, " consent shall be considered obtained whenever one party has announced to all other parties" that the conversation is about to be recorded, provided the announcement is itself recorded. Connecticut accepts consent obtained in writing or on the recording, a verbal notice recorded at the beginning, or " an automatic tone warning device". In Montana, a warning takes a recording outside the offence and " If one person provides the warning, either party may record". In Illinois and Massachusetts the offence is secret recording: Illinois defines surreptitious as " obtained or made by stealth or deception, or executed through secrecy or concealment", and Massachusetts defines interception as to " secretly hear, secretly record". For an archive, the practical point is the same everywhere: the notice has to be on each recording, because a buyer asks for proof per batch, not a policy that says notices were played.

Florida and Oregon split by channel. Florida's all-party rule covers calls, but for in-person speech it protects only a speaker " exhibiting an expectation that such communication is not subject to interception". Oregon is the reverse of what most people expect: a participant may record a phone call, but an in-person conversation may be recorded only if " all participants in the conversation are specifically informed". Meeting audio recorded in an Oregon office is held to a stricter rule than the same company's calls.

Nevada, Delaware and Michigan rest on readings. Nevada's call statute lets a person intercept a call only with " the prior consent of one of the parties" and an emergency in which a court order is impractical, which one party's consent alone does not meet; for in-person talk, one participant's authorization is enough. Delaware's privacy offence names interception " without the consent of all parties", while its wiretap chapter allows a participant to record, so the safe reading is all-party. Michigan's felony names " the consent of all parties", but defines eavesdropping as recording " the private discourse of others", which can be read to leave a participant's own recording outside it. The labels for Nevada and Michigan follow the statute text; the court rulings that have read these sections were not read for this guide.

A call has two ends. A call between a California customer and an agent in Texas involves both states' rules, and an archive rarely records where each caller sat. The careful reading for an archive is the strictest state either end could have been in, which in practice means sorting calls by the caller's area code or account address before a buyer sees them, and treating any batch that includes all-party states as needing all-party consent.

The gap: consent to record is not consent to train

Every statute in the table governs the act of recording, and several govern using or disclosing a recording that was unlawfully made. None of them, as read for this guide, says whether a recording that was lawfully made, with every consent its state required at the time, may later be licensed to train a model without asking the speakers again. That is the question holders most need answered, and no statute we read answers it. Fiund, which licenses call archives, puts the same point plainly: " A recording notice and consent to training use are different questions". California's own definition shows why: a recording notice can make a call non-confidential under Cal. Penal Code 632(c), but it speaks to the expectation of being recorded, not to what the recording may later be used for.

Two states do say something about later use, for particular kinds of recordings, and both bear directly on call archives.

Illinois forbids passing on telemarketing and opinion-research recordings. Illinois lets a business engaged in marketing, opinion research or telephone solicitation record calls on one party's consent for quality control, staff training and internal research. But nothing obtained under that exemption may be " divulged to any third party", in the statute's closing words. Licensing such recordings to an AI buyer divulges them to a third party, so they need the all-party basis instead.

Pennsylvania limits customer-service recordings to the business's own training, and to one year. Its exception lets a telemarketing or customer-service business record calls on one party's consent, but " Any communications recorded pursuant to this paragraph may only be used by the business for the purpose of training or quality control", and, unless other law requires otherwise, they "shall be destroyed within one year from the date of recording". Licensing them to an outside buyer is use by someone other than the business, and the statute does not say whether "training" means only training staff. Unless another law requires keeping them, recordings older than a year that rest on this exception should already have been destroyed.

The defect travels to the buyer. Three statutes reach not only the person who recorded unlawfully but whoever later receives or uses the recording. The federal Act punishes using or disclosing contents known to be unlawfully intercepted. Oregon bars obtaining a recording " while knowing or having good reason to believe" it was unlawfully made, and using or divulging it. Illinois makes it eavesdropping to use or disclose information one knows or reasonably should know was obtained in violation of the Article. This is why buyers ask for consent records batch by batch, and why a well-advised buyer will not take an archive whose consent basis nobody can show.

Other law reaches recordings too. A voice can be biometric data: Illinois lists a " voiceprint" among biometric identifiers, and forbids any private entity to " sell, lease, trade, or otherwise profit from" one. A recording is not on that list by itself, but a voiceprint extracted from one is the kind of identifier the statute covers, which is why cautious buyers want written releases naming AI training and sale and a promise not to extract identifiers. In California, a services contract clause allowing a digital replica of the individual's voice in place of their work is unenforceable unless it describes the uses with reasonable specificity, so old talent agreements with presenters and voice actors should not be stretched to cover a voice model. And software has terms of its own: Zoom's API terms forbid using customer content obtained through its APIs to train, develop, evaluate, test, improve, or modify artificial intelligence or machine learning models, so how meeting recordings leave a platform matters as much as whether they may be sold.

Outside the US the question has a clearer answer. For holders in Canada, the federal privacy law says that when personal information " is to be used for a purpose not previously identified", the individual's consent is required before it is used that way, and the Privacy Commissioner's guidance on call recording says an organisation " may only use the information collected for the purposes specified". For holders in the UK, the Information Commissioner's guidance on purpose limitation says a very different or unexpected purpose is likely to be incompatible, and that " you're likely to need to ask for specific consent". Licensing identifiable call recordings for AI training is the kind of new purpose that guidance describes.

What follows for a holder is practical rather than doctrinal. Among the rights questions a buyer's counsel asks are whether calls and meetings were recorded with every participant's consent, and whether that consent covered use beyond the meeting. The answer a careful buyer looks for is recording notices on every call, written consent from internal participants for AI licensing, and external participants excluded or separately consented. The valuation on this site follows the same line: it does not value speech unless every speaker agreed to this use. Where the speakers did not, there are two routes: ask them (customers and staff can be asked, and a consented slice is a sensible first sale), or sell the words without the voices, which is the subject of the next section.

De-identifying voices and transcripts

Text can be de-identified well; voices cannot yet. That difference decides how most call and meeting archives can be sold.

A voice identifies its speaker, and disguising it does not reliably hold. Voice anonymisation (changing a voice so a listener or a machine cannot tell who spoke) is an active research field, and its own benchmark is sobering: the organisers of the VoicePrivacy 2024 Challenge reported that top-performing systems which first appeared to hide speakers well were later shown to be vulnerable to stronger attacks. For that reason our de-identification standard does not sell recordings of voices unless the people in them consented in writing to AI training and sale; transcripts are rewritten instead. Several buyers draw the same line from the other side: micro1's licence terms have the holder warrant that the data contains no biometric identifiers or biometric information, card or bank numbers, or children's data, unless the order form says otherwise, and HUMXN says that " Names, faces, voices, account numbers and other personal data are removed or masked" before buyers see a dataset.

Spoken personal details can be cut from audio that is sold. Where every speaker consented and the audio itself is the product, what remains is the content: names, account numbers, card numbers and addresses spoken aloud. The catalogue datasets of real support calls in the price table are sold PII-redacted, and Fiund's process catalogues identifying detail and then redacts or excludes it, with "material that fails the consent review" not licensed at all.

Transcripts are where most archives can be sold. A transcript carries what a call or meeting was about (the task, the problem, the resolution, the way people actually talk) without the voice. Rewritten so that no person can be identified, it is a text record, and text records are what many buyers of company data ask for: Handshake's outbound team licenses companies' code and support transcripts among their operating data. The de-identification engine behind our standard was built for exactly this kind of record:

• It runs on your computer. The engine works in your browser, so nothing identifying leaves your computer: not the transcripts, not the list of people it builds, not the key.
• It protects a person in every mention. A speaker counts as protected only when every mention of them is replaced, with each person's name, first name, surname, address and handle linked to one stand-in.
• It screens content as well as names. Calls about health, children, beliefs or crime are flagged under its content screen, because removing names does not make a health conversation safe to sell.
• A person reads a sample. Rules read every record, and a person reads a random sample of the rewritten ones before anything goes further.
• The buyer commits in writing. Every licence binds the buyer not to re-identify the data or link it to other data to do so.

On company Slack, email and documents the engine scores 86.8% on a public benchmark, and we call its output de-identified, with the measured risk, not anonymous. You can run it yourself, free, with the de-identification tool. When you add files to the valuation, your browser reads them: recordings are counted, their words are read from transcripts (VTT and SRT files), and what leaves your computer is counts, with no text and no names of people.

What buyers' terms ask of audio

A buyer's page says what it will pay for; its terms say what it will not take. For audio, the terms gather around five questions: whose voices these are and whether they agreed, whether the holder owns the recordings, what personal details they contain, how they were recorded, and what the buyer may do with them afterwards.

Three contradictions in those terms are worth knowing before you rely on any of them.

A conversation with people you know is someone else's audio. Luel pays for conversations recorded " with two people you know", while its terms have the contributor warrant that the content includes no third-party audio. Two other people in a recording are third parties; a contributor who records them is relying on their consent, and the terms put that risk on the contributor.

The highest rate buys the cheapest licence. The Voices.com job's rate is the highest in this research, but it buys evaluation rights, and evaluation licences fetch about 0.2 to 0.5 times a full licence. A holder comparing offers should compare what is granted before comparing the number.

"No standard rate card" is the standard. Fiund, Murf, Protege and Troveo publish no rate for existing audio, and Murf's page says only that compensation "depends on the dataset's language, quality, and length". That leaves the holder with the market figures in this guide and with competing offers as the only references.

The deeper clauses (warranties, indemnities, exclusivity, term, audits) are the same for audio as for any data, and our guide to what a data licensing agreement contains covers them. Two are sharper for audio. Exclusivity has a price: an exclusive licence fetches about 2.5 to 5 times a non-exclusive one, and non-exclusive data can be sold again, which an archive of common speech is well placed to do. And withdrawal: where a buyer's terms, like Kled's, require a refund when a speaker withdraws consent, an archive with many speakers carries a refund risk for every one of them.

How to prepare an archive for sale

Preparing an audio archive is mostly sorting it by consent, then making the part that can be sold as useful to a buyer as possible. A buyer's lawyers ask 27 rights questions of a data sale, and 16 of them stop it until answered; for audio, most of the blockers are about consent and personal details. What follows is the order to do it in.

1. Inventory what you hold

List hours, number of recordings, date ranges, languages, and the recording setup: phone system or meeting software, mono or dual-channel, sample rate, file format, and whether originals or compressed copies were kept. Fiund's list of what buyers ask for starts here: hours and call counts, consent records, personal details and their redaction status, "Channel separation", and call metadata such as "intents, outcomes, queue types, languages, and dates". Note which recordings already have transcripts and speaker labels, because those are worth more than audio alone.

2. Map consent, batch by batch

For each batch, record where the speakers were (the caller's state, or the meeting's office), what the recording notice said and whether it is on the recording, and what your privacy notice and customer terms said about use. Then sort into four lots:

• Every speaker agreed to AI use. Recordings made under consent that names AI training, staff who signed a consent covering it, contributors who signed releases. This audio can be sold as audio.
• Lawfully recorded, no AI consent. Most call and meeting archives. Sell transcripts de-identified, or ask the speakers; a consented slice of recent calls is often enough for a first sale.
• Doubtful or defective. Recordings from all-party states without proof of consent, Illinois telemarketing recordings made under the one-party exemption, Pennsylvania customer-service recordings made under its exception, anything under legal hold. Leave these out.
• Not yours to grant. A contact centre or outsourcer recording its clients' customers: Fiund's advice is that "Your client agreements decide what you can grant", so the client's contract is the first document to read.

3. Clear everything else in the recording

For podcasts, radio and broadcast archives, consent is only one layer. Guests and co-hosts have rights in their voices; licensed music beds, jingles and syndicated segments belong to others, and music rights holders are the most active litigants in AI (our video guide sets out the music suits). Advertisements read on air may be owned by the advertiser. Talent agreements written before AI rarely cover a voice model, and in California a vague one may not be enforceable for digital replica uses at all. Deliver speech without music where you can, and list what you removed.

4. Remove personal details

From transcripts, with the de-identification tool, in your browser. From audio that will be sold as audio, by silencing or replacing the spans with names, numbers and addresses, and say in the datasheet which method was used. Health, children and payment card conversations are better left out than redacted.

5. Describe it

A buyer cannot listen to thousands of hours. A datasheet that states hours by language, accent and channel, speaker counts, recording conditions, transcript accuracy, consent basis per lot and what was removed is what lets it say yes quickly; the datasheet tool structures one. The guide on training data provenance explains why buyers now ask for this evidence up front, and the audio and speech page lists what raises and lowers the price of speech.

6. Know your number, then let buyers compete

Value the archive before any buyer names a price: the valuation prices conversational speech at $25 to $150 an hour of speech, adjusted for the kind of recording and the language and tapered for large archives, and on its starting answers a mid-sized archive of two-way English conversation in which every speaker agreed comes out at $5,000 to $30,000. It binds no one and needs no upload. Then compare offers on the grant, the exclusivity and the withdrawal terms as well as the price: the offer check places a written offer against the market and lists that buyer's clauses.

Example: a regional insurer's contact centre. Take a company with several years of recorded customer calls, mono recordings from its phone system, a "calls may be recorded for quality and training" message at the start of each call, customers across several states including California and Pennsylvania, and no transcripts. Its first finding is that the audio as it stands sits at the bottom of the market: bulk call audio is listed at well under a dollar an hour, and even redacted, transcribed, two-channel support calls list at $24.50 an hour. Its second is consent: its notice supports the recording, but no notice mentioned AI, Pennsylvania's exception limits customer-service recordings to the business's own use and one year, California's mobile-phone rule needs every party, and insurance calls carry health and payment details. Its practical route is the transcript: transcribe the recent calls, de-identify them, drop health and payment conversations, and offer the text as records of how claims and complaints are actually resolved, with the audio held back. If buyers want audio, a new notice that names AI training, played and recorded on every call from now on, builds a consented archive for the next sale.

Example: a podcast network. Take a network with a decade of interview podcasts in English and Spanish, hosts on staff, guests who signed a simple appearance release, and music beds in every episode. Its rights are in better shape than a call centre's (everyone knew they were being recorded, and the hosts are its employees) but its audio is the most plentiful kind there is, which is why general podcast audio lists at under five dollars an hour. What lifts it: the original multitrack files with each speaker on its own channel, transcripts, Spanish and any accented or regional speech, and stripping the music. What blocks part of it: guest releases that do not mention AI, which a careful buyer will want refreshed or will discount, and voice cloning, which Protege, Murf and the Twine project all rule out. Exclusivity is worth keeping off the table, because a general archive earns more licensed several times than once.

How a sale through us works

AIData.Broker acts as the holder's agent in the sale: we work only for the holder, and the holder itself licenses or sells its recordings to the buyer. We run sales only for holders based in the United States, the United Kingdom and Canada, and we take on a lot only where the expected sale reaches $26,000; smaller archives are usually better served by listing directly with a collector or catalogue, and the page on going direct or through us sets out the trade-off.

The steps are the same as for any lot, set out in how it works:

1. A valuation first. The valuation puts a range on the archive from public prices, binds no one, and needs no upload. For speech it asks four questions: hours, kind of speech, language, and whether every speaker agreed.
2. The rights file. Together we sort the archive by consent as in the section above, decide what is sold as audio and what as de-identified transcripts, and write the datasheet.
3. Every buyer that fits, at once. The lot goes first as a blind profile, without the holder's name, to every buyer whose needs and terms fit and that we can screen, and all of them answer by the same deadline. The holder may leave out any buyer it chooses.
4. Offers in writing, side by side. Every offer reaches the holder with its full terms as the buyer stated them, what the holder would keep after our fee, and anything that buyer pays whoever introduces a holder. The holder alone decides whether to accept, which buyer and on what terms.
5. Delivery and payment between holder and buyer. The holder, or a provider it chooses, delivers to the buyer, and the buyer pays the holder directly.

Our fee is a success fee on what the holder is actually paid, band by band: 15% of the first $250,000, 10% from $250,000 to $1 million, and 6% above $1 million, as the pricing page sets out, so on a sale paying the holder $300,000 the fee is $42,500. Nothing is payable if nothing sells. In a sale we run we take nothing from any buyer: where a buyer offers an introduction fee (Shaip, for one, pays one to whoever brings it a holder), we decline it, tell the holder the amount and ask the buyer to put it into the holder's price instead. The mandate is exclusive for 2 months, with a tail of 12 months for sales to buyers we introduced; the binding rules are in the mandate, which our terms describe.

Questions people ask

How much is an hour of speech worth for AI training? It depends on the side of the market. Licensed speech sold for a median of $76 per hour of audio per hour, ready-made datasets ask $118 per hour of audio, and speech recorded to order costs $122 per hour of audio. Existing archive audio sits far lower: the one published offer is up to Rs. 200 per audio hour.

Can a company sell its call recordings to AI companies? Sometimes, and more often as de-identified transcripts than as audio. The recordings must have been lawfully made under the strictest state involved, the consent must cover the new use or the speakers must be asked, and personal details must come out. Illinois forbids divulging telemarketing recordings made on one party's consent, and Pennsylvania limits customer-service recordings to the business's own training or quality control.

Is "this call may be recorded" enough consent to sell the recording? No statute we read says it is, and none says it is not. A notice can satisfy the recording laws (in Washington a recorded announcement counts as everyone's consent), but it speaks to being recorded, not to later use, and in Canada and the UK a new purpose generally needs new consent. Buyers treat it as a gap and price it.

Which states require everyone's consent to record a call? From the statutes read for this guide: California, Florida, Maryland, New Hampshire, Pennsylvania and Washington for calls and conversations alike; Connecticut for calls, by its civil statute; Nevada for calls, in effect; Massachusetts and Illinois for secret recording; Delaware and Michigan are disputed; Montana requires the knowledge of all parties. The table above links each to its text.

Can I sell my podcast archive for AI training? Yes, if you hold the rights to the hosts' and guests' voices and remove music you licensed in. Shaip offers up to Rs. 200 per audio hour for Indic-language podcasts under a training-only licence; general English podcast audio lists at about $4.86 an hour. Original multitrack files, transcripts and less common languages raise the price.

Can voices be anonymised before sale? Not reliably. Voice anonymisation systems that first looked strong were later broken by stronger attacks, so we sell recordings of voices only where every speaker consented in writing to AI training and sale, and rewrite transcripts instead. The de-identification tool does that in your browser.

How much do people get paid to record their voice for AI? A median of $24 per hour of audio per approved hour of audio across published rates, from $7.50 to $32 an hour for phone sessions, and $100 per finished hour for studio voice work. Read the licence: some projects promise no voice likeness, and some buy evaluation rights only.

What this means for you

Selling recorded speech is a consent exercise with a price attached. The holders who earn the most per hour can show that every speaker agreed to this use, deliver original two-channel audio with transcripts, and offer speech that is scarce; everyone else earns from transcripts, or from a consented slice, or from the next archive they record with the right notice. The steps below turn that into a plan.

1. Know your number before anyone quotes one. The rates page and the price board hold every public price per hour of audio, and the valuation puts a range on your own archive from them. The audio and speech page keeps the current prices, buyers and 11 lots of speech on offer.
2. Sort by consent first. Map each batch to the states its speakers were in and to what they were told, using the state table, and split the archive into what can be sold as audio, what as de-identified transcripts, and what stays out.
3. Sell the words where you cannot sell the voices. Rewrite transcripts with the de-identification tool and read the de-identification standard for what buyers will rely on.
4. Read the grant before the rate. Training or evaluation, exclusive or not, perpetual or for a term, and what happens if a speaker withdraws: the offer check lists each buyer's clauses, and the guides on exclusive and non-exclusive licences and when AI data buyers pay cover the terms that decide what you keep.
5. Fix the archive you are still recording. A notice that names AI training, played and recorded on every call, and a release that names it for every guest, turn tomorrow's recordings into the consented kind buyers pay most for.
6. Match the route to the size. A small or general archive is usually best placed directly with a collector or catalogue; a large, distinctive or consented archive is where competing offers make a difference. How it works describes a sale through us, and the rights hub and the hub on what labs buy cover the rest of the market.

AIData.Broker acts as the holder's agent: the holder itself licenses or sells its data to the buyer, and we do not take ownership of it. The holder pays us a fee fixed in the mandate and paid only if a sale closes, and we run sales only for holders based in the United States, the United Kingdom and Canada.