Speech Data Collection Cost and Prices per Hour (2026)

There is no single price for an hour of speech data: what an hour costs an AI company depends far more on how it was made than on what was said. Across the public prices, the median hour recorded to a buyer's order costs $122 per hour of audio, an hour bought off the shelf from a corpus or a catalogue asks $119 per hour of audio, and an hour in sales that actually happened went for $73 per hour of audio. The people who record it are paid a median of $25 per hour of audio for an approved hour, so buyers pay about 5.0x what speakers receive. The open programmes that publish their full budgets, which pay for recruiting, transcription and checking as well as the recording, spent $244 to $333 per hour of audio.

This guide is for the people who budget for speech data and the people who hold it: an AI team deciding whether to order new recordings or buy a corpus, a procurement or finance lead checking a vendor's quote, and a company with call recordings, podcasts or broadcasts wondering what they are worth beside newly recorded speech. It covers the five prices of an hour of speech, what collecting speech to order costs and what is in that cost, what speakers are paid, what corpora and catalogues charge, how language changes the price, what has actually been paid, where AI companies buy speech, and when to build rather than buy. Every figure links to its source on the price board; the audio and speech page keeps the current anchors.

AIData.Broker 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, not legal, tax or investment advice.

Contents

1. The five prices of an hour of speech
2. What collecting speech to order costs
3. What speakers are paid
4. Speech datasets for sale: corpora and catalogues
5. How language moves the price
6. What has actually been paid
7. Where AI companies buy speech training data
8. Build or buy: recording new speech or licensing it
9. What raises and lowers the price of an hour
10. Questions people ask
11. What this means for you

1. The five prices of an hour of speech

Speech is priced by the hour of audio almost everywhere in this market, and that makes it look like one commodity with one price. It is not. The same hour carries a different price at each point it passes through: what a speaker is paid to record it, what a vendor charges to record it to a buyer's specification, what a corpus or catalogue asks for a copy of it, what a buyer has actually paid for it in a completed sale, and what it cost a programme to produce it end to end. Comparing a quote with the wrong one of these is the most common budgeting mistake, and the reason the medians below differ several times over.

The table sets the five side by side. Each median is computed from every public price of its kind on the price board, converted to a price per hour of audio, and it moves as new prices are added.

Two of these surprise people. The first is that off the shelf is not cheaper than to order at the median: a corpus is a finished, checked, documented product with a licence attached, and the catalogues price it that way. The second is that completed sales sit below both asking prices. Asking prices are opening positions; the sales that happened include large bulk licences, older corpora and regional markets, and they are what a buyer can actually expect to pay. The rates page shows how each of these medians is built and how many prices sit behind it.

The middle half of all public prices for audio, whatever their side, runs $65 to $266 per hour of audio. Anything quoted far outside that band is either unusual speech (studio voice data, a rare language over the telephone) or bulk audio sold with few guarantees.

2. What collecting speech to order costs

Collecting speech to order means a vendor recruits speakers who fit the buyer's specification (languages, accents, ages, devices, settings), records them reading prompts or holding conversations, transcribes and checks the audio, and delivers it with the consent the buyer needs. It is the only route when the speech a buyer wants does not exist yet: a dialect no corpus covers, a device or noise condition the product will meet, a conversation type no archive holds, or speech with consent written for training from the start.

The vendors that publish a price quote it per delivered hour, and their ranges are wide because the specification moves the cost more than anything else.

Public tenders give the same picture from the buyer's side, with the price a contract actually awarded. A Chinese government procurement for a multilingual corpus priced each hour of personalised recorded speech at CNY 960 an hour for 750 hours and each hour of targeted annotated speech at CNY 650 an hour for 1,125 hours, both awarded to Lifewood. Taiwan's Hakka Affairs Council awarded its corpus collection and system for 2026 to 2028 at NT$19,600,000, and a Chinese broadcaster paid CNY 186,000 to collect one city's dialect. A Singapore research agency split a collection of Malay and Tagalog speech between two suppliers, at SGD 120,000 and SGD 100,000.

What is inside the price

A vendor's per-hour price bundles four costs, and knowing them lets a buyer read a quote. The speaker's pay is the smallest part (section 3). Transcription is the largest single line for conversational speech: human transcription of English is sold at $1.99 /min by Rev, $1.30 per minute by Speechpad and Starting at $0.79/min by TranscribeMe, so transcribing one hour costs more than recording it. Then recruiting and managing speakers to hit a demographic mix, and checking every recording for quality, consent and the specification.

The programmes that publish their whole budgets show what all of it adds up to when nothing is cut. The African Next Voices programme reportedly set aside $2.2 million to record 9,000 hours across 18 languages. NaijaVoices says creating its dataset of 1,800+ hours in Hausa, Igbo and Yoruba cost substantial (>$600k). India's Bhashini mission reportedly contributed $5-$6 Mn to the IndicVoices collection, which released 7,348 hours. These are the costs of building speech data from nothing in languages with few existing recordings, and they set an upper bound a commercial quote rarely exceeds.

3. What speakers are paid

The people who record speech for AI are paid per approved hour of audio, per session or per hour of their time, and the rates are public because programmes recruit openly. They matter to buyers for two reasons: they are the floor under every price above, and a programme that pays too little gets speech that is rushed, repetitive or recorded by the same few people.

Three patterns run through the table. Conversation pays more than reading, because it needs two people, a scenario and more checking. Professional voice talent is a different market from crowd recording: a finished studio hour pays several times a crowd hour, and it is what speech-synthesis buyers want. And a low-resource language does not always pay its speakers more: Luel's rate for low-resource conversation is below its English rate, even though such speech is scarcer for buyers. The scarcity shows up in what buyers pay, not always in what speakers are paid. Across advertised roles, the median pay for voice acting and voice-over work in AI training is $35 an hour (live median), as tracked by AITraining.jobs, a site under common ownership with AIData.Broker.

4. Speech datasets for sale: corpora and catalogues

Most speech bought off the shelf comes from three kinds of source. Corpus distributors such as the Linguistic Data Consortium (LDC) and ELRA license research and commercial corpora built over decades, many of them the benchmarks speech recognition was measured on. Data vendors' catalogues, such as SpeechData.ai, Defined.ai and Nexdata, sell ready-made datasets recorded for training. Listing sites such as Opendatabay and Datarade carry datasets from many suppliers, including bulk call-centre and podcast audio priced by the dataset. The lots register lists 14 lots of audio on offer across these sites today.

Their prices differ by how a dataset was made and what the licence allows.

Read per hour, the same table spans three orders of magnitude, and the chart shows it on one scale.

Sources: Hindi call-centre bulk, Hindi podcasts, AISHELL-1, Fisher English, Hindi service calls, US service calls, ALLIES, Hindi conversations, French conversations, Defined.ai plan, US English mobile.

Three things explain most of the spread. The licence: a corpus licensed for research only lists at a median of $15 per hour of audio an hour, and the same corpus for commercial use costs several times that; the LDC's prices above are its non-member fees, and its members pay differently. Age and fame: Fisher and AISHELL-1 are old, widely used corpora that every speech team already knows, so their per-hour price is low; a lab gains little from training on what every competitor has. How the audio was gathered and what comes with it: bulk call-centre and podcast audio sold by the hundreds of thousands of hours, with few guarantees on consent or transcripts, sits at the bottom, while prompted speech from known, consenting speakers, recorded on the devices a product will meet and transcribed, sits at the top.

5. How language moves the price

Language matters, but less simply than buyers expect. Within one catalogue and one kind of recording, a rarer language usually costs more per hour; across catalogues, the format, the licence and the age of the corpus swamp it. The table keeps to commercial catalogue prices so the comparison is as like for like as the public record allows.

Per hour, the colloquial Egyptian Arabic telephone corpus is the dearest speech on the list and the Tamil read corpus one of the cheapest, though both are in languages with far fewer recordings than English. The difference is the recording: telephone speech from many speakers in a colloquial dialect is hard to gather, while a short read-speech corpus from a research project is cheap to reproduce. The Hausa, Igbo and Yoruba suite shows the other effect of language: an open project that wants its data used asks a token sum for commercial use, which puts a floor of near zero under what a lab will pay a commercial supplier for the same languages.

Prices for scarce languages also move with demand. One report on Odia, a low-resource Indian language, says an hour of its speech used to cost about $3-$4, now it's $40 as Indian AI developers began buying. A holder of speech in a language that is now in demand holds something whose price has risen, and the buyers that publish where they take speech from are on the buyer board.

6. What has actually been paid

Asking prices are what a listing hopes for. The sales that happened are fewer and mostly come from public filings and procurement records, but they are what anchors the median of $73 per hour of audio.

The largest public record of speech data sales comes from China, where Beijing Haitian Ruisheng Science Technology (SpeechOcean), a listed speech data company, reports the hours it sold and its revenue. In 2020 its off-the-shelf speech recognition datasets sold at an average of 413 元/小时 (CNY 413 an hour, average over 198,951 hours sold), and its speech synthesis datasets at 13,832 元/小时 (CNY 13,832 an hour, average over 1,029 hours sold). The gap between the two is the clearest number on record for how much more studio voice data is worth than recognition data: synthesis needs hours of one consistent, professional voice, recognition needs many ordinary voices. The same Chinese procurement that bought recordings to order also bought finished, ready-made speech data at CNY 500 an hour for 1,125 hours, below the price it paid for new recordings.

At the other end of the market, buyers of existing archives offer little per hour. Shaip tells holders of Indic-language podcasts, radio and broadcasts that it pays up to Rs. 200 per audio hour for audio they already have, a small fraction of what newly recorded speech costs. That gap is the subject of the guide to selling audio and speech, which explains why archive audio sells cheapest and what raises its price: consent to train, transcripts, speaker separation and a language buyers want.

A different model pays by use rather than by the hour. In ElevenLabs' voice library, a person who shares a cloned voice earns a royalty when others generate speech with it, by default $0.03 per 1,000 characters generated. It prices a voice, not a recording, and it is the model a holder of voice talent recordings may meet.

7. Where AI companies buy speech training data

AI companies get speech in five ways, and most use several at once. Which one fits depends on whether the speech they need already exists, how much they need and how sure they must be of its consent.

1. Corpus distributors. The LDC and ELRA license the corpora speech research was built on, for research or commercial use, at published prices. They are the first stop for a benchmark or a well-known language, and the last for anything new.
2. Data vendors' catalogues and collection services. Firms such as Appen, Shaip, Defined.ai, Nexdata, SpeechData.ai and AIxBlock sell ready-made datasets and record new speech to order. This is where most commercial speech comes from.
3. Listing sites. Opendatabay and Datarade carry datasets from many suppliers, from bulk call-centre audio to curated conversation sets, each with its own asking price; the lots register reads them side by side.
4. Recording programmes that pay people directly. Apps and programmes such as Perit AI and Luel recruit speakers and pay them per approved hour (section 3), selling what they record to model developers.
5. Holders of existing audio. Call centres, podcasters, broadcasters and companies with years of recorded meetings license what they already hold, usually through an intermediary that collects, cleans and licenses it on. The guide to selling audio and speech covers who buys it and what it needs before it can be sold.

A buyer that wants a specific kind of speech found and licensed, rather than shopped for itself, can also ask a broker to source it: AIData.Broker's page for buyers who want data sourced describes how that works on our side.

8. Build or buy: recording new speech or licensing it

The choice between recording new speech and licensing existing speech is mostly a choice about time, consent and fit, and only partly about price. The medians put the two close together ( $122 per hour of audio to order against $119 per hour of audio off the shelf), so the decision usually turns on the other three.

Consent is the line that most often decides it. Speech recorded under a two-party consent law was consented to for recording, and that consent is not automatically consent to train a model: California, for one, requires the consent of all parties to record any call with a cell or cordless phone on one end. Licensed archives from call centres or meetings need that gap closed, through de-identification or fresh consent, before a careful buyer takes them; the audio guide's consent section goes through it state by state. Speech recorded to order avoids the problem by writing training into the consent form, which is a large part of what its higher price buys.

9. What raises and lowers the price of an hour

Whether you are pricing a quote or a holding, the same features move an hour up or down.

What raises it:

• Conversation over reading, and two speakers on separate channels over one mixed track.
• Accurate human transcripts, time-aligned and speaker-labelled; transcription alone is priced per minute (section 2).
• Consent written for training, documented per speaker.
• A language or accent buyers are short of, recorded in the conditions a product will meet: telephones, cars, noisy rooms.
• Many distinct speakers with documented demographics, rather than many hours from a few.
• Studio voice data for synthesis, which sold at many times the price of recognition data in the one public record of both ( Haitian Ruisheng's 2020 sales).
• Exclusivity, priced separately; the guide to exclusive and non-exclusive licences explains how.

What lowers it:

• Bulk audio with no transcripts or unclear consent.
• Speech already public or already in widely sold corpora.
• A research-only licence, which lists at a median of $15 per hour of audio an hour.
• Personal details left in (names, account numbers, health information), which a buyer must pay to remove or will not take at all; the de-identification standard sets out what has to go.

Questions people ask

How much does it cost to collect speech data? Vendors that record speech to a buyer's order charge a median of $122 per hour of audio per hour of delivered audio, with published ranges from $50 to $300 per hour for general collection to $200 to $500 per hour for studio voice data. Open programmes that publish their full budgets spent $244 to $333 per hour of audio, including recruiting, transcription and checking.

How much are people paid to record speech for AI? A median of $25 per hour of audio per approved hour of audio. Conversation pays more than read speech (Perit pays $28 / hr for conversation against $20 per approved audio hour for reading), and professional voice talent is paid per finished hour, up to $120-400 USD per hour.

Where can I buy speech datasets? From corpus distributors (the LDC and ELRA), from data vendors' catalogues (speechdata.ai, Defined.ai, Nexdata and others), and from listing sites such as Opendatabay and Datarade. The lots register lists 14 lots of audio on offer across these sites, each with its asking price.

Is it cheaper to buy a speech corpus than to record new speech? Not always. At the median, off-the-shelf speech asks $119 per hour of audio an hour and speech recorded to order costs $122 per hour of audio. Old research corpora are cheap per hour, and bulk call-centre audio cheaper still, but new speech recorded to a specification comes with consent written for training and fits the product.

How much does transcription add to the cost of speech data? Human transcription of English is sold at $1.30 per minute to $1.99 /min, so transcribing an hour of conversation can cost more than recording it. Machine transcripts checked by people cost less, and buyers price accurate, speaker-labelled transcripts into the hour.

Why do low-resource languages cost more? Because few recordings exist and they are hard to gather, so buyers compete for them: one report says an hour of Odia speech used to cost about $3-$4, now it's $40. But the format often matters more than the language: a short read-speech corpus in a rare language can be cheaper per hour than colloquial telephone speech in a common one.

What is my company's call or meeting audio worth? Less per hour than newly recorded speech, because it was recorded for another purpose and its consent rarely covers training; one buyer offers holders of Indic archive audio up to Rs. 200 per audio hour. Transcripts, separated speakers, consent and a language buyers want raise it. The guide to selling audio and speech and the valuation put a figure on a specific archive.

What this means for you

Speech is one of the few kinds of training data with prices public at every point from the speaker to the completed sale, and that makes a quote easy to check. The useful habit is to know which of the five prices a number is, and to compare it only with the same kind.

1. Check a quote against its own side of the market. A collection quote goes against $122 per hour of audio, a catalogue price against $119 per hour of audio, a completed licence against $73 per hour of audio. The rates page and the price board hold every price behind them.
2. Price the transcript and the consent, not just the audio. They are where most of the cost and most of the risk sit.
3. Buy the old corpora for benchmarks, and record or license what is new for training. What every competitor already has adds little.
4. If you hold speech, know what it is before you are offered for it. The audio guide covers who buys archives and what they need, the valuation gives a range from these prices, and the offer check reads a written offer against the market.
5. If you need speech found, say what you need. The buyer board lists who sells and buys speech, and the acquire page describes how we source data for buyers.

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.