Skip to content
Voice & Speech Data

Speech Data Collection Cost per Hour: 10 Drivers to Price In Before You Buy

Share:
Speech Data Collection Cost per Hour: 10 Drivers to Price In Before You Buy

Last updat­ed: Novem­ber 2026

Key takeaways

  • No sin­gle pub­lic rate exists. Speech data col­lec­tion cost per hour depends on ten dri­vers, led by lan­guage rar­i­ty, speech style, tran­scrip­tion depth and licens­ing rights.
  • Define the hour first. A price per record­ed hour and a price per accept­ed hour are not com­pa­ra­ble until you write down accep­tance cri­te­ria.
  • Tran­scrip­tion can out­weigh record­ing. In one 2025 study of Bam­bara, accu­rate tran­scrip­tion took about 30 hours of human labor per audio hour in the lab and 36 in the field.
  • Off-the-shelf is cheap­er up front, not always cheap­er over­all. Check com­mer­cial rights, chan­nel fit and speak­er meta­da­ta before you buy.
  • Use the HOURS spec check (Hours, Own­er­ship, Users, Record­ing, Stan­dards) to get quotes you can com­pare line by line.

Ask three ven­dors for their speech data col­lec­tion cost per hour and you may get three num­bers that can­not be com­pared. One prices record­ed hours, anoth­er prices accept­ed hours, and a third leaves out tran­scrip­tion or com­mer­cial rights.

This guide is for ASR, TTS and con­ver­sa­tion­al AI teams in the US and EU who need to bud­get a cus­tom voice dataset. You get the 10 cost dri­vers that move a quote and a clear def­i­n­i­tion of accept­ed hours. You also get an off-the-shelf vs cus­tom com­par­i­son, licens­ing mod­els and the HOURS spec check­list for com­par­ing ven­dors.

Want to judge qual­i­ty before you talk price? Request a free 30-minute speech sam­ple set record­ed and tran­scribed to your spec.

How much does speech data collection cost per hour?

Speech data col­lec­tion cost per hour has no sin­gle pub­lic rate. Ten dri­vers set it: lan­guage rar­i­ty, speak­ers and quo­tas, speech style, chan­nel and envi­ron­ment, accept­ed-hour rules, tran­scrip­tion depth, QA depth, con­sent and meta­da­ta, licens­ing rights and turn­around. Com­mon-lan­guage read speech sits at the low end, and exclu­sive con­ver­sa­tion­al data in a rare accent at the high end.

Few neu­tral sources pub­lish per-hour prices for cus­tom col­lec­tion. Most ven­dors ask for a spec before they quote, and that is fair. Two projects with the same hour count can need very dif­fer­ent amounts of work.

What pub­lic, non-ven­dor sources do offer are ref­er­ence points. As of Novem­ber 2026, these are the most use­ful ones we could ver­i­fy:

  • Label­ing at scale. The Peo­ple’s Speech paper (NeurIPS 2021) says the team “ini­tial­ly esti­mat­ed that hand-label­ing 100,000 hours of speech would cost $5,000,000.” That works out to about $50 per hour, for label­ing only. It is a 2021 plan­ning esti­mate, not a mar­ket rate.
  • Automa­tion changes the math. The same team instead ran forced align­ment on 52,500 hours of exist­ing audio with tran­scripts. It took three days on Google Cloud at an esti­mat­ed cost of $3,000. That only works when tran­scripts already exist.
  • Human effort per hour. In a 2025 field study of Bam­bara tran­scrip­tion, the team cor­rect­ed machine drafts of 53 hours of speech. That took about 30 hours of human labor per audio hour in lab con­di­tions, and 36 in the field. Effort varies by lan­guage and lit­er­a­cy, but the ratio shows why tran­scrip­tion dom­i­nates many bud­gets.
  • Licensed cor­po­ra. Under the Lin­guis­tic Data Con­sor­tium’s licens­ing terms, non­mem­ber licensees, includ­ing for-prof­it non­mem­bers, can­not use LDC data for any com­mer­cial pur­pose. Com­mer­cial rights come through for-prof­it mem­ber­ship. Rights, not just hours, set the price.

In short: treat any sin­gle per-hour fig­ure with cau­tion. Price your own spec, dri­ver by dri­ver, and ask ven­dors to show their line items.

Why do speech data quotes vary so much?

Quotes vary because ven­dors price dif­fer­ent units and bun­dle dif­fer­ent work. One quote cov­ers record­ed hours with machine tran­scripts. Anoth­er cov­ers accept­ed hours with ver­ba­tim tran­scripts, a sec­ond QA pass and exclu­sive rights. Until two quotes describe the same spec, the same accep­tance rules and the same license, their per-hour num­bers are not com­pa­ra­ble.

Voice dataset pric­ing uses at least five units in the mar­ket. Know­ing which one you are look­ing at is the first step in a fair com­par­i­son.

  • Per record­ed hour. All cap­tured audio, includ­ing silence, retakes and files that may lat­er fail checks.
  • Per accept­ed hour. Only audio that pass­es your writ­ten accep­tance cri­te­ria. The rate looks high­er, but rejects are the ven­dor’s cost.
  • Per speak­er or per ses­sion. Com­mon when many speak­ers each con­tribute a few min­utes, such as accent cov­er­age projects.
  • Per utter­ance or prompt. Typ­i­cal for wake words and short com­mands, where hours are a poor mea­sure.
  • Per audio minute tran­scribed. Used when you already have record­ings and only need tran­scrip­tion or anno­ta­tion.

Ask every ven­dor to split the quote into line items: recruit­ment and con­sent, record­ing, tran­scrip­tion, anno­ta­tion, QA, licens­ing pre­mi­um and project man­age­ment. Line items make hid­den gaps vis­i­ble.

In short: nor­mal­ize the unit before you com­pare the num­ber.

What are the 10 cost drivers of speech data collection?

The ten dri­vers are lan­guage and accent rar­i­ty, speak­ers and demo­graph­ic quo­tas, speech style, chan­nel and record­ing envi­ron­ment, vol­ume and accept­ed-hour rules, tran­scrip­tion depth, QA depth, con­sent and meta­da­ta, licens­ing rights and turn­around. Lan­guage, speech style, tran­scrip­tion and licens­ing usu­al­ly swing a quote the most. Turn­around main­ly mat­ters when you need a rush deliv­ery.

Cost dri­verLow­er-effort endHigh­er-effort endTyp­i­cal swing
1. Lan­guage and accent rar­i­tyWide­ly spo­ken lan­guage, gen­er­al accentLow-resource lan­guage or named region­al accentsHigh
2. Speak­ers and quo­tasFew­er speak­ers, loose tar­getsMany unique speak­ers, hard age, gen­der and region quo­tasMedi­um
3. Speech styleScript­ed read promptsSpon­ta­neous mono­logue or two-par­ty con­ver­sa­tionHigh
4. Chan­nel and envi­ron­mentSmart­phone in a qui­et roomTele­pho­ny pairs, stu­dio cap­ture, far-field or set noise con­di­tionsMedi­um
5. Vol­ume and accept­ed-hour rulesRaw record­ed hours, loose rulesNet speech hours under strict accep­tance cri­te­riaMedi­um
6. Tran­scrip­tion depthOrtho­graph­ic text, no time­stampsVer­ba­tim with tags, word time­stamps and diariza­tionHigh
7. QA depthSam­ple-based audit100% sec­ond pass, dou­ble tran­scrip­tion, WER auditsMedi­um
8. Con­sent, meta­da­ta and com­pli­anceSin­gle-pur­pose con­sent, basic meta­da­taPer-pur­pose con­sent, rich meta­da­ta, EU data han­dlingMedi­um
9. Licens­ing and usage rightsNon-exclu­sive licenseExclu­sive rights or full own­er­shipHigh
10. Turn­aroundFlex­i­ble sched­uleRush deliv­ery with par­al­lel teamsLow
Speech data cost dri­vers and the typ­i­cal swing each one caus­es. Swing is our qual­i­ta­tive read, not a mea­sured fig­ure.
Ten speech data cost drivers rated by relative cost impact: language, speakers, style, channel, volume, transcription, QA, consent, licensing and turnaround
The 10 cost dri­vers and how far each can typ­i­cal­ly swing a speech data quote (qual­i­ta­tive).

1. Language and accent rarity

Recruit­ing is the main bot­tle­neck, since speak­ers of a wide­ly spo­ken lan­guage are easy to find. A low-resource lan­guage or a spe­cif­ic region­al accent needs local recruiters, native tran­scribers and extra screen­ing to con­firm each speak­er real­ly has that accent.

Spec­i­fy: lan­guage, locale, accent or dialect labels, and how speak­er claims are ver­i­fied. For India-spe­cif­ic plan­ning, see our Indic speech data col­lec­tion guide.

2. Speakers and demographic quotas

More unique speak­ers per hour means more recruit­ing, onboard­ing and con­sent records per hour. Hard quo­tas also slow the end of a project, because the last few cells, such as speak­ers over 60 in one region, take the longest to fill.

Spec­i­fy: total speak­ers, a cap on min­utes per speak­er, quo­ta cells, and which quo­tas are hard and which are best effort.

3. Speech style: scripted, spontaneous or conversational

Script­ed speech is the fastest to record and to tran­scribe, because the prompt is a start­ing tran­script. Spon­ta­neous speech needs top­ic prompts and full tran­scrip­tion. Two-par­ty con­ver­sa­tion adds sched­ul­ing, a sep­a­rate chan­nel per speak­er and turn-lev­el tran­scrip­tion with over­laps.

If you are build­ing a phone or app agent, read what voice agent train­ing data needs before you choose the mix.

4. Channel and recording environment

Tele­pho­ny audio is nar­row­band, usu­al­ly sam­pled at 8 kHz. Wide­band app audio is often 16 kHz, and TTS voic­es are usu­al­ly record­ed at 44.1 or 48 kHz in treat­ed rooms. Each set­up needs dif­fer­ent devices, rooms and checks.

Real envi­ron­ments such as cars, kitchens or streets mean record­ing peo­ple where they are, then tag­ging each file. Con­trolled noise con­di­tions and far-field micro­phone arrays add book­ing and equip­ment time. Nev­er accept upsam­pled nar­row­band audio labeled as wide­band.

5. Volume and the accepted-hour definition

Larg­er vol­umes spread fixed costs, such as guide­lines, tool­ing and pilot work, over more hours. But the def­i­n­i­tion of an hour often moves the price more than the vol­ume does. We cov­er this in the next sec­tion.

6. Transcription depth

An ortho­graph­ic tran­script records the words, while ver­ba­tim also cap­tures fillers, false starts and cut-off words. Each extra lay­er, such as noise and over­lap tags, seg­ment or word time­stamps, speak­er diariza­tion and lan­guage tags for code-switch­ing, adds a pass and a guide­line to fol­low. That is why tran­scrip­tion cost per hour of audio climbs fast.

Tran­scrip­tion appears here only as a cost dri­ver. If you need it on its own, our human audio tran­scrip­tion ser­vices page lists the ver­ba­tim and time-aligned options.

7. QA depth and WER audits

Sam­ple-based audits check a share of each batch. Eval­u­a­tion sets often need 100% sec­ond-pass review or dou­ble tran­scrip­tion. A WER audit has a review­er re-tran­scribe a sam­ple, then mea­sures the word error rate between the deliv­ered tran­script and that ref­er­ence.

Spec­i­fy: the audit sam­ple size, the met­ric, the pass thresh­old per batch and what hap­pens to batch­es that fail.

Per-pur­pose con­sent, a con­sent ID on every file and rich meta­da­ta (age band, gen­der, region, device, envi­ron­ment) all take time to cap­ture and check. That work pro­tects you lat­er, when you need to prove cov­er­age or hon­or a dele­tion request.

If any speak­ers are in the EU, the GDPR applies to their record­ings. Per the Euro­pean Com­mis­sion, bio­met­ric data processed sole­ly to iden­ti­fy a per­son is sen­si­tive data with extra con­di­tions. Our speak­er con­sent form tem­plate cov­ers the details. This is gen­er­al infor­ma­tion, not legal advice.

9. Licensing exclusivity and usage rights

A non-exclu­sive license is cheap­est, because the ven­dor can license the same audio to oth­ers. Exclu­sive rights or full own­er­ship remove that resale val­ue, so the price ris­es. Rights for syn­thet­ic voice or voice cloning need sep­a­rate speak­er con­sent, which also affects cost.

10. Turnaround

Rush time­lines mean par­al­lel recruit­ing, more review­ers and tighter QA win­dows. Turn­around rarely changes the per-hour price as much as the dri­vers above, but it can force trade-offs in quo­ta bal­ance. Build pilot time into your plan.

In short: four dri­vers do most of the work. Pin down lan­guage, speech style, tran­scrip­tion depth and rights first, and the rest of the quote gets much eas­i­er to com­pare.

What is the difference between recorded and accepted hours?

Record­ed hours count all cap­tured audio, includ­ing silence, retakes and files that fail checks. Accept­ed hours count only audio that meets your writ­ten accep­tance cri­te­ria, such as sam­ple rate, noise lev­el, speech con­tent, tran­script accu­ra­cy and com­plete meta­da­ta. Pric­ing per accept­ed hour moves the risk of rejects to the ven­dor. The rate looks high­er, but it buys usable data.

It helps to sep­a­rate four kinds of “hour” in every con­tract:

  1. Ses­sion hours. Time speak­ers spend in a record­ing ses­sion, includ­ing set­up, instruc­tions and breaks.
  2. Record­ed hours. Total audio cap­tured, includ­ing silence and retakes.
  3. Net speech hours. Audio left after trim­ming long silences and remov­ing retakes.
  4. Accept­ed hours. Net speech that pass­es every accep­tance check and is approved by you.

The gap between raw and usable hours is real, even out­side cus­tom col­lec­tion. Peo­ple’s Speech start­ed with 52,500 hours of audio with tran­scripts. After forced align­ment and fil­ter­ing out mis­matched tran­scripts, 31,400 hours were aligned, about 60% of the source.

Bar chart from the People's Speech paper: 52,500 source hours produced 31,400 aligned hours, about 60 percent usable
Raw hours shrink after qual­i­ty fil­ter­ing: 52,500 source hours became 31,400 aligned hours in Peo­ple’s Speech (Galvez et al., 2021).

Your yield will dif­fer, but the les­son holds. Write accep­tance cri­te­ria before record­ing starts, so both sides price the same thing:

  • Audio for­mat: sam­ple rate, bit depth, chan­nel count, file type, no clip­ping.
  • Speech con­tent: min­i­mum speech ratio and silence trim­ming rules.
  • Envi­ron­ment: noise lev­el or envi­ron­ment tag match­es the spec.
  • Tran­script accu­ra­cy: max­i­mum WER on an audit­ed sam­ple.
  • Meta­da­ta: every required field filled, plus a valid con­sent ID.
  • Quo­tas: speak­er caps and quo­ta cells respect­ed per batch.
  • Rework: rejec­tion rea­sons, rework rules and dead­lines.

In short: a cheap record­ed hour can be an expen­sive accept­ed hour. Com­pare speech data col­lec­tion cost per hour on accept­ed hours, not record­ed ones.

How is a speech data quote built?

A reli­able quote fol­lows five steps: a writ­ten spec, a free sam­ple against that spec, a small paid pilot, scaled batch deliv­ery and approval of each batch. The pilot is where the per-hour price becomes evi­dence-based, because it mea­sures real reject rates, tran­scrip­tion effort and how fast quo­ta cells fill.

Five-step speech data quote process: spec, sample, pilot, scale and approve, with accepted hours growing at each step
How a reli­able speech data quote is built, from writ­ten spec to approved hours.
  1. Spec. Your team and the ven­dor’s project lead agree lan­guages, speak­ers, style, audio for­mat, tran­scrip­tion guide­line, accep­tance cri­te­ria and license. Out­put: a signed spec.
  2. Sam­ple. The ven­dor records and tran­scribes a short sam­ple to that spec. You check audio, tran­scripts and meta­da­ta against your own cri­te­ria.
  3. Pilot. A small paid batch runs end to end. You mea­sure reject rate, rework time and quo­ta fill, then con­firm or adjust the price.
  4. Scale. Batch­es ship with quo­ta reports and QA reports. Native-speak­er review­ers audit each batch before deliv­ery.
  5. Approve. You accept or reject each batch against the spec. In pay-on-approval mod­els, only approved files are invoiced.

In short: do not sign a large order off a rate card. Let a pilot set the price.

Is off-the-shelf speech data cheaper than custom collection?

Off-the-shelf speech data is usu­al­ly cheap­er and faster up front, because record­ing costs are shared across buy­ers. Cus­tom col­lec­tion costs more per hour but match­es your accents, chan­nel, domain and rights. The cheap­er option is the one that fits. A free cor­pus that does not match pro­duc­tion audio, or for­bids com­mer­cial use, can cost more in the end.

Fac­torOff-the-shelf or open dataCus­tom col­lec­tion
Upfront costLow, some­times freeHigh­er, priced to your spec
Time to first dataFast, often imme­di­ateLonger, after a sam­ple and pilot
Chan­nel and domain fitWhat­ev­er the cor­pus cap­turedMatched to your prod­uct
Speak­er meta­da­taOften lim­it­edFields you define
Com­mer­cial rightsVaries: open, research-only or licensedWrit­ten into your con­tract
Exclu­siv­i­tyNone, oth­ers use the same dataAvail­able at a pre­mi­um
Off-the-shelf vs cus­tom speech data at a glance.

Open data can be a strong start. Peo­ple’s Speech offers about 30,000 hours of most­ly Eng­lish speech under CC-BY-SA, with a CC-BY sub­set, and its paper describes Mozil­la Com­mon Voice as CC0-licensed. Both suit pre­train­ing and research, but check whether their accents, chan­nels and domains match your users.

Licens­ing his­to­ry mat­ters too. The same paper recounts that Baidu col­lect­ed 12,000 hours of read speech from 9,600 speak­ers through crowd­sourc­ing. A legal review lat­er found the speak­er agree­ments did not per­mit pub­lic release or unre­strict­ed com­mer­cial reuse, so the data was nev­er released.

Most teams end up with a hybrid that uses open or licensed data for breadth and cus­tom data for their own domain and chan­nel. They also keep a cus­tom held-out test set with speak­ers who nev­er appear in train­ing.

In short: buy breadth, col­lect fit, and always own your test set.

Which speech data licensing model fits your budget?

Pick the nar­row­est speech data licens­ing mod­el that cov­ers your real use. Open licens­es cost noth­ing but car­ry attri­bu­tion or share-alike terms. Research-only licens­es are cheap but block prod­ucts. Non-exclu­sive com­mer­cial licens­es suit most teams. Exclu­sive licens­es and full own­er­ship cost the most, so reserve them for data that gives you a last­ing edge.

Licens­ing mod­elWhat you getCost effect
Open license (CC0, CC-BY, CC-BY-SA)Free use under the license terms, such as attri­bu­tion or share-alikeLow­est
Research-only licenseNon­com­mer­cial research and edu­ca­tion useLow, but not usable in prod­ucts
Non-exclu­sive com­mer­cial licenseUse in prod­ucts; the ven­dor may license the same data to oth­ersMod­er­ate
Time- or field-lim­it­ed exclu­siveNo resale to oth­ers for a set peri­od or fieldHigh­er
Full own­er­ship or assign­mentYou own the dataset out­rightHigh­est
Com­mon speech data licens­ing mod­els and their rel­a­tive cost.

What­ev­er the mod­el, list the rights you need in writ­ing:

  • Train­ing, fine-tun­ing and eval­u­a­tion of your mod­els.
  • TTS or syn­thet­ic voice use, backed by sep­a­rate speak­er con­sent.
  • Use by affil­i­ates, con­trac­tors and cloud proces­sors.
  • Deriv­a­tive datasets, such as aug­ment­ed or fil­tered ver­sions.
  • Term, ter­ri­to­ry and what hap­pens to data and mod­els if the license ends.

This sec­tion is gen­er­al infor­ma­tion, not legal advice. Have coun­sel review any license before you sign.

The HOURS spec check: how to get comparable speech data quotes

We use a sim­ple tool to turn the ten dri­vers into a spec any ven­dor can price: the HOURS spec check. Each let­ter cov­ers two dri­vers, and if any let­ter is blank, the quote you get back is a guess.

The HOURS spec check framework for speech data quotes: Hours, Ownership, Users, Recording and Standards, each covering two cost drivers
The HOURS spec check groups the 10 cost dri­vers into five let­ters you can fill in before ask­ing for quotes.
Let­terCov­ers dri­versQues­tions to answer
H: HoursVol­ume and accept­ed-hour rules; turn­aroundHow many accept­ed hours? What counts as accept­ed? When do batch­es ship?
O: Own­er­shipLicens­ing rights; con­sent and meta­da­taExclu­sive or not? Which uses? Which con­sent pur­pos­es and meta­da­ta fields?
U: UsersLan­guage and accent; speak­ers and quo­tasWhich lan­guages and accents? How many speak­ers? Which quo­tas are hard?
R: Record­ingSpeech style; chan­nel and envi­ron­mentScript­ed, spon­ta­neous or con­ver­sa­tion­al? Which sam­ple rate, devices and set­tings?
S: Stan­dardsTran­scrip­tion depth; QA depthWhich tran­scrip­tion guide­line and tags? Which QA met­ric and thresh­old?
The HOURS spec check maps each of the 10 cost dri­vers to a ques­tion you can answer before request­ing quotes.

HOURS checklist: spec fields to fill in

Copy these fields into your request for quote. Exam­ple entries show the lev­el of detail that gets a pre­cise answer.

Spec fieldExam­ple entry
Use caseASR fine-tun­ing for a cus­tomer sup­port voice agent
Lan­guages and localesUS Eng­lish; Ital­ian (Italy)
Accents or dialectsUS South­ern, US Mid­west; North­ern and South­ern Ital­ian
Accept­ed hours200 accept­ed hours per lan­guage, net of silence
Speak­ers and capsAt least 500 speak­ers per lan­guage, 30 min­utes max­i­mum each
Quo­tasAge bands 18–30, 31–50, 51+; gen­der with­in an agreed tol­er­ance
Speech style mix20% script­ed, 30% spon­ta­neous, 50% two-par­ty role-play
Chan­nel and for­mat8 kHz tele­pho­ny pairs, one speak­er per chan­nel, 16-bit WAV
Envi­ron­mentsQui­et room, home with TV, car, street
Tran­scrip­tionVer­ba­tim, noise and over­lap tags, seg­ment time­stamps, speak­er turns
QA and accep­tanceAudit­ed sam­ple per batch, WER thresh­old agreed after pilot
Meta­da­ta and con­sentSpeak­er ID, age band, gen­der, region, device, envi­ron­ment, con­sent ID
LicenseNon-exclu­sive com­mer­cial license for train­ing and eval­u­a­tion
Time­lineSam­ple, pilot, then week­ly batch­es
The HOURS check­list spec fields with illus­tra­tive exam­ple entries. Replace every entry with your own require­ments.

Send your spec, get a scoped quote. Fill in the HOURS fields and send them to our voice data team. We will scope a pilot billed only on approved files. Request the HOURS check­list (men­tion “HOURS check­list” in your mes­sage and we’ll email the one-page PDF).

Common mistakes that inflate speech data costs

Most over­runs in the cost of audio data col­lec­tion come from deci­sions made before the first file is record­ed. These are the ones we see most often:

  • Com­par­ing dif­fer­ent hours. A record­ed-hour quote and an accept­ed-hour quote look alike on paper but buy dif­fer­ent amounts of usable audio.
  • Writ­ing accep­tance rules at deliv­ery. If cri­te­ria arrive late, reject­ed batch­es become dis­putes instead of rework.
  • Record­ing the wrong chan­nel. Stu­dio audio for a phone prod­uct, or phone audio for a TTS voice, means pay­ing twice.
  • Over-tran­scrib­ing train­ing data. Bulk train­ing audio may only need ortho­graph­ic text. Save ver­ba­tim, tags and dou­ble review for the eval­u­a­tion set.
  • Loose quo­tas. Vague demo­graph­ic tar­gets skew toward the eas­i­est speak­ers to recruit, and re-col­lect­ing the gaps costs more lat­er.
  • Buy­ing research-only data for a prod­uct. A low price means lit­tle if the license for­bids com­mer­cial use.
  • Pay­ing for exclu­siv­i­ty you do not need. Exclu­sive rights make sense for dif­fer­en­ti­at­ing data, not for gener­ic read speech.
  • Skip­ping the pilot. With­out a pilot, nobody knows the real reject rate, so both sides guess and pad the price.

In short: the cheap­est hour is the one you only pay for once.

How Graveiens AI quotes speech data collection

Graveiens AI does not pub­lish a rate card, because your spec sets your speech data col­lec­tion cost per hour. We quote against your writ­ten spec, bro­ken into line items you can com­pare with any oth­er ven­dor.

  • Free 30-minute speech sam­ple set. Record­ed and tran­scribed to your spec, so you can check qual­i­ty before a pilot.
  • Pilot billed only on approved files. You are invoiced only for deliv­er­ables you approve, at the pilot stage and after.
  • Explic­it-con­sent work­flow. Par­tic­i­pant onboard­ing, then record­ing man­age­ment, then meta­da­ta tag­ging, then QA, with a con­sent record behind every file.
  • Lan­guage cov­er­age. 25+ lan­guages, includ­ing Indi­an lan­guages and Euro­pean lan­guages such as Ital­ian.
  • Qual­i­ty sys­tem. Graveiens AI is ISO 9001 cer­ti­fied.

Script­ed, spon­ta­neous and con­ver­sa­tion­al col­lec­tion are all avail­able through our con­sent-first voice and speech data ser­vices, and mul­ti-modal pro­grams run through our cus­tom data col­lec­tion pro­grams. For any­thing out­side a stan­dard spec, we scope it with you.

We do not promise mod­el accu­ra­cy gains, because those depend on your archi­tec­ture and eval­u­a­tion. What we com­mit to is data that match­es the spec you sign off.

Frequently asked questions

What is a typical speech data collection cost per hour?

There is no reli­able typ­i­cal fig­ure, because the spec dri­ves the price. Pub­lic ref­er­ence points exist: the Peo­ple’s Speech team esti­mat­ed hand-label­ing 100,000 hours at $5 mil­lion in 2021, about $50 per hour for label­ing alone. Treat such fig­ures as con­text, and price your own spec with a sam­ple and a pilot.

What is an accepted hour in speech data collection?

An accept­ed hour is an hour of net speech audio that pass­es every accep­tance check in your spec. Checks cov­er audio for­mat, noise lev­el, tran­script accu­ra­cy, meta­da­ta com­plete­ness and speak­er caps. It excludes silence, retakes and reject­ed files. Pric­ing per accept­ed hour puts the cost of rejects on the ven­dor.

Is transcription included in speech data collection pricing?

Some­times, and that is a com­mon source of con­fu­sion. Some quotes cov­er record­ing only, some include machine drafts, and some include human ver­ba­tim tran­scrip­tion with time­stamps and tags. Ask for tran­scrip­tion as a sep­a­rate line item, with the guide­line and QA thresh­old named, so you can com­pare quotes fair­ly.

Why does conversational speech data cost more than read speech?

Con­ver­sa­tion­al data needs two speak­ers sched­uled togeth­er, a sep­a­rate chan­nel for each, and sce­nar­ios that pro­duce nat­ur­al turn-tak­ing. Tran­scrip­tion is hard­er too, with over­laps, fillers and speak­er turns to mark. Read speech starts from a known script, so both record­ing and tran­scrip­tion take less effort per hour.

Does exclusive licensing increase the cost of speech data?

Yes. With a non-exclu­sive license, the ven­dor can license the same audio to oth­er buy­ers, which spreads the cost. Exclu­sive rights or full own­er­ship remove that option, so the price ris­es. Reserve exclu­siv­i­ty for data that gives you a last­ing advan­tage, such as your own domain or a rare accent.

How can I lower speech data collection cost without hurting quality?

Define accept­ed hours and accep­tance cri­te­ria up front, match the chan­nel to your prod­uct, and tran­scribe bulk train­ing audio at a lighter lev­el than your eval­u­a­tion set. Use open or licensed data for breadth, run a pilot before scal­ing, and only pay for exclu­siv­i­ty where it mat­ters.

Is open speech data free to use commercially?

Not always. Licens­es dif­fer. CC0 and CC-BY gen­er­al­ly allow com­mer­cial use, CC-BY-SA adds share-alike terms, and research-only licens­es block com­mer­cial use. LDC, for exam­ple, lim­its non­mem­ber licensees to non­com­mer­cial research. Read the license for each cor­pus and have coun­sel con­firm your intend­ed use.

Get a free 30-minute speech sample set

Tell us your lan­guage, accent, speech style and chan­nel. We will record and tran­scribe a free 30-minute sam­ple set to your spec, so you can check it against your own accep­tance cri­te­ria. If you go ahead, the pilot is billed only on approved files.

Request your free speech sam­ple set

Request the HOURS check­list (men­tion “HOURS check­list” in your mes­sage and we’ll email the one-page PDF).

Sources

  1. Galvez, D. et al. The Peo­ple’s Speech: A Large-Scale Diverse Eng­lish Speech Recog­ni­tion Dataset for Com­mer­cial Usage. NeurIPS 2021 Datasets and Bench­marks Track, arXiv:2111.09344. https://arxiv.org/abs/2111.09344
  2. Diar­ra, Y., Coulibaly, N. S. and Lev­en­thal, M. Cost Analy­sis of Human-cor­rect­ed Tran­scrip­tion for Pre­dom­i­nate­ly Oral Lan­guages. arXiv:2510.12781, Octo­ber 2025. https://arxiv.org/abs/2510.12781
  3. Lin­guis­tic Data Con­sor­tium, Uni­ver­si­ty of Penn­syl­va­nia. Licens­ing. https://www.ldc.upenn.edu/data-management/using/licensing
  4. Euro­pean Com­mis­sion. What per­son­al data is con­sid­ered sen­si­tive? https://commission.europa.eu/law/law-topic/data-protection/rules-business-and-organisations/legal-grounds-processing-data/sensitive-data/what-personal-data-considered-sensitive_en

About the author: Graveiens AI Edi­to­r­i­al Team. The Graveiens AI edi­to­r­i­al team writes about speech data, AI train­ing data and data com­pli­ance for AI builders in the US and EU. The team works along­side our voice col­lec­tion, tran­scrip­tion and QA leads, so each guide reflects how real projects are scoped and deliv­ered.

Jitendra Choubay
Jitendra Choubay
CEO & Founder

Jitendra Choubay is the CEO & Founder of Graveiens AI, leading a human-in-the-loop data services team that helps AI builders with data collection, annotation, consent-backed voice data, transcription and LLM fine-tuning. He writes on building better, ethically sourced AI training data.

Get the next Graveiens AI article

Expert notes on AI data, annotation, LLMs and eLearning — no spam, unsubscribe anytime.

Need AI Development? Data Annotation? eLearning?

Talk to Graveiens AI about data collection, annotation, voice data, RLHF/SFT, LLM evaluation and AI training data — invoiced only on approved work.

Contact Graveiens AI