{"id":396,"date":"2026-11-19T18:30:00","date_gmt":"2026-11-19T13:00:00","guid":{"rendered":"https:\/\/www.graveiensai.com\/blog\/?p=396"},"modified":"2026-11-12T00:15:03","modified_gmt":"2026-11-11T18:45:03","slug":"speech-data-collection-cost","status":"publish","type":"post","link":"https:\/\/www.graveiensai.com\/blog\/speech-data-collection-cost\/","title":{"rendered":"Speech Data Collection Cost per Hour: 10 Drivers to Price In Before You Buy"},"content":{"rendered":"\n<p class=\"last-updated wp-block-paragraph\"><em>Last updat\u00aded: Novem\u00adber 2026<\/em><\/p>\n\n\n<div class=\"wp-block-group key-takeaways\"><div class=\"wp-block-group__inner-container is-layout-flow wp-block-group-is-layout-flow\">\n<h2 class=\"wp-block-heading\" id=\"key-takeaways\">Key takeaways<\/h2>\n\n\n<ul class=\"wp-block-list\"><li><strong>No sin\u00adgle pub\u00adlic rate exists.<\/strong> Speech data col\u00adlec\u00adtion cost per hour depends on ten dri\u00advers, led by lan\u00adguage rar\u00adi\u00adty, speech style, tran\u00adscrip\u00adtion depth and licens\u00ading rights.<\/li><li><strong>Define the hour first.<\/strong> A price per record\u00aded hour and a price per accept\u00aded hour are not com\u00adpa\u00adra\u00adble until you write down accep\u00adtance cri\u00adte\u00adria.<\/li><li><strong>Tran\u00adscrip\u00adtion can out\u00adweigh record\u00ading.<\/strong> In one 2025 study of Bam\u00adbara, accu\u00adrate tran\u00adscrip\u00adtion took about 30 hours of human labor per audio hour in the lab and 36 in the field.<\/li><li><strong>Off-the-shelf is cheap\u00ader up front, not always cheap\u00ader over\u00adall.<\/strong> Check com\u00admer\u00adcial rights, chan\u00adnel fit and speak\u00ader meta\u00adda\u00adta before you buy.<\/li><li><strong>Use the HOURS spec check<\/strong> (Hours, Own\u00ader\u00adship, Users, Record\u00ading, Stan\u00addards) to get quotes you can com\u00adpare line by line.<\/li><\/ul>\n\n<\/div><\/div>\n\n\n<p class=\"wp-block-paragraph\">Ask three ven\u00addors for their speech data col\u00adlec\u00adtion cost per hour and you may get three num\u00adbers that can\u00adnot be com\u00adpared. One prices record\u00aded hours, anoth\u00ader prices accept\u00aded hours, and a third leaves out tran\u00adscrip\u00adtion or com\u00admer\u00adcial rights.<\/p>\n\n\n<p class=\"wp-block-paragraph\">This guide is for ASR, TTS and con\u00adver\u00adsa\u00adtion\u00adal AI teams in the US and EU who need to bud\u00adget a cus\u00adtom voice dataset. You get the 10 cost dri\u00advers that move a quote and a clear def\u00adi\u00adn\u00adi\u00adtion of accept\u00aded hours. You also get an off-the-shelf vs cus\u00adtom com\u00adpar\u00adi\u00adson, licens\u00ading mod\u00adels and the HOURS spec check\u00adlist for com\u00adpar\u00ading ven\u00addors.<\/p>\n\n\n<p class=\"cta-line wp-block-paragraph\"><strong>Want to judge qual\u00adi\u00adty before you talk price?<\/strong> <a href=\"https:\/\/www.graveiensai.com\/contact-us\">Request a free 30-minute speech sam\u00adple set<\/a> record\u00aded and tran\u00adscribed to your spec.<\/p>\n\n\n<h2 class=\"wp-block-heading\" id=\"how-much-does-speech-data-collection-cost-per-hour\">How much does speech data collection cost per hour?<\/h2>\n\n\n<p class=\"wp-block-paragraph\">Speech data col\u00adlec\u00adtion cost per hour has no sin\u00adgle pub\u00adlic rate. Ten dri\u00advers set it: lan\u00adguage rar\u00adi\u00adty, speak\u00aders and quo\u00adtas, speech style, chan\u00adnel and envi\u00adron\u00adment, accept\u00aded-hour rules, tran\u00adscrip\u00adtion depth, QA depth, con\u00adsent and meta\u00adda\u00adta, licens\u00ading rights and turn\u00adaround. Com\u00admon-lan\u00adguage read speech sits at the low end, and exclu\u00adsive con\u00adver\u00adsa\u00adtion\u00adal data in a rare accent at the high end.<\/p>\n\n\n<p class=\"wp-block-paragraph\">Few neu\u00adtral sources pub\u00adlish per-hour prices for cus\u00adtom col\u00adlec\u00adtion. Most ven\u00addors ask for a spec before they quote, and that is fair. Two projects with the same hour count can need very dif\u00adfer\u00adent amounts of work.<\/p>\n\n\n<p class=\"wp-block-paragraph\">What pub\u00adlic, non-ven\u00addor sources do offer are ref\u00ader\u00adence points. As of Novem\u00adber 2026, these are the most use\u00adful ones we could ver\u00adi\u00adfy:<\/p>\n\n\n<ul class=\"wp-block-list\"><li><strong>Label\u00ading at scale.<\/strong> <a href=\"https:\/\/arxiv.org\/abs\/2111.09344\" target=\"_blank\" rel=\"noopener\">The Peo\u00adple\u2019s Speech paper (NeurIPS 2021)<\/a> says the team \u201cini\u00adtial\u00adly esti\u00admat\u00aded that hand-label\u00ading 100,000 hours of speech would cost $5,000,000.\u201d That works out to about $50 per hour, for label\u00ading only. It is a 2021 plan\u00adning esti\u00admate, not a mar\u00adket rate.<\/li><li><strong>Automa\u00adtion changes the math.<\/strong> The same team instead ran forced align\u00adment on 52,500 hours of exist\u00ading audio with tran\u00adscripts. It took three days on Google Cloud at an esti\u00admat\u00aded cost of $3,000. That only works when tran\u00adscripts already exist.<\/li><li><strong>Human effort per hour.<\/strong> In <a href=\"https:\/\/arxiv.org\/abs\/2510.12781\" target=\"_blank\" rel=\"noopener\">a 2025 field study of Bam\u00adbara tran\u00adscrip\u00adtion<\/a>, the team cor\u00adrect\u00aded machine drafts of 53 hours of speech. That took about 30 hours of human labor per audio hour in lab con\u00addi\u00adtions, and 36 in the field. Effort varies by lan\u00adguage and lit\u00ader\u00ada\u00adcy, but the ratio shows why tran\u00adscrip\u00adtion dom\u00adi\u00adnates many bud\u00adgets.<\/li><li><strong>Licensed cor\u00adpo\u00adra.<\/strong> Under <a href=\"https:\/\/www.ldc.upenn.edu\/data-management\/using\/licensing\" target=\"_blank\" rel=\"noopener\">the Lin\u00adguis\u00adtic Data Con\u00adsor\u00adtium\u2019s licens\u00ading terms<\/a>, non\u00admem\u00adber licensees, includ\u00ading for-prof\u00adit non\u00admem\u00adbers, can\u00adnot use LDC data for any com\u00admer\u00adcial pur\u00adpose. Com\u00admer\u00adcial rights come through for-prof\u00adit mem\u00adber\u00adship. Rights, not just hours, set the price.<\/li><\/ul>\n\n\n<p class=\"wp-block-paragraph\"><strong>In short:<\/strong> treat any sin\u00adgle per-hour fig\u00adure with cau\u00adtion. Price your own spec, dri\u00adver by dri\u00adver, and ask ven\u00addors to show their line items.<\/p>\n\n\n<h2 class=\"wp-block-heading\" id=\"why-do-speech-data-quotes-vary-so-much\">Why do speech data quotes vary so much?<\/h2>\n\n\n<p class=\"wp-block-paragraph\">Quotes vary because ven\u00addors price dif\u00adfer\u00adent units and bun\u00addle dif\u00adfer\u00adent work. One quote cov\u00aders record\u00aded hours with machine tran\u00adscripts. Anoth\u00ader cov\u00aders accept\u00aded hours with ver\u00adba\u00adtim tran\u00adscripts, a sec\u00adond QA pass and exclu\u00adsive rights. Until two quotes describe the same spec, the same accep\u00adtance rules and the same license, their per-hour num\u00adbers are not com\u00adpa\u00adra\u00adble.<\/p>\n\n\n<p class=\"wp-block-paragraph\">Voice dataset pric\u00ading uses at least five units in the mar\u00adket. Know\u00ading which one you are look\u00ading at is the first step in a fair com\u00adpar\u00adi\u00adson.<\/p>\n\n\n<ul class=\"wp-block-list\"><li><strong>Per record\u00aded hour.<\/strong> All cap\u00adtured audio, includ\u00ading silence, retakes and files that may lat\u00ader fail checks.<\/li><li><strong>Per accept\u00aded hour.<\/strong> Only audio that pass\u00ades your writ\u00adten accep\u00adtance cri\u00adte\u00adria. The rate looks high\u00ader, but rejects are the ven\u00addor\u2019s cost.<\/li><li><strong>Per speak\u00ader or per ses\u00adsion.<\/strong> Com\u00admon when many speak\u00aders each con\u00adtribute a few min\u00adutes, such as accent cov\u00ader\u00adage projects.<\/li><li><strong>Per utter\u00adance or prompt.<\/strong> Typ\u00adi\u00adcal for wake words and short com\u00admands, where hours are a poor mea\u00adsure.<\/li><li><strong>Per audio minute tran\u00adscribed.<\/strong> Used when you already have record\u00adings and only need tran\u00adscrip\u00adtion or anno\u00adta\u00adtion.<\/li><\/ul>\n\n\n<p class=\"wp-block-paragraph\">Ask every ven\u00addor to split the quote into line items: recruit\u00adment and con\u00adsent, record\u00ading, tran\u00adscrip\u00adtion, anno\u00adta\u00adtion, QA, licens\u00ading pre\u00admi\u00adum and project man\u00adage\u00adment. Line items make hid\u00adden gaps vis\u00adi\u00adble.<\/p>\n\n\n<p class=\"wp-block-paragraph\"><strong>In short:<\/strong> nor\u00admal\u00adize the unit before you com\u00adpare the num\u00adber.<\/p>\n\n\n<h2 class=\"wp-block-heading\" id=\"what-are-the-10-cost-drivers-of-speech-data-collection\">What are the 10 cost drivers of speech data collection?<\/h2>\n\n\n<p class=\"wp-block-paragraph\">The ten dri\u00advers are lan\u00adguage and accent rar\u00adi\u00adty, speak\u00aders and demo\u00adgraph\u00adic quo\u00adtas, speech style, chan\u00adnel and record\u00ading envi\u00adron\u00adment, vol\u00adume and accept\u00aded-hour rules, tran\u00adscrip\u00adtion depth, QA depth, con\u00adsent and meta\u00adda\u00adta, licens\u00ading rights and turn\u00adaround. Lan\u00adguage, speech style, tran\u00adscrip\u00adtion and licens\u00ading usu\u00adal\u00adly swing a quote the most. Turn\u00adaround main\u00adly mat\u00adters when you need a rush deliv\u00adery.<\/p>\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>Cost dri\u00adver<\/th><th>Low\u00ader-effort end<\/th><th>High\u00ader-effort end<\/th><th>Typ\u00adi\u00adcal swing<\/th><\/tr><\/thead><tbody><tr><td>1. Lan\u00adguage and accent rar\u00adi\u00adty<\/td><td>Wide\u00adly spo\u00adken lan\u00adguage, gen\u00ader\u00adal accent<\/td><td>Low-resource lan\u00adguage or named region\u00adal accents<\/td><td>High<\/td><\/tr><tr><td>2. Speak\u00aders and quo\u00adtas<\/td><td>Few\u00ader speak\u00aders, loose tar\u00adgets<\/td><td>Many unique speak\u00aders, hard age, gen\u00adder and region quo\u00adtas<\/td><td>Medi\u00adum<\/td><\/tr><tr><td>3. Speech style<\/td><td>Script\u00aded read prompts<\/td><td>Spon\u00adta\u00adneous mono\u00adlogue or two-par\u00adty con\u00adver\u00adsa\u00adtion<\/td><td>High<\/td><\/tr><tr><td>4. Chan\u00adnel and envi\u00adron\u00adment<\/td><td>Smart\u00adphone in a qui\u00adet room<\/td><td>Tele\u00adpho\u00adny pairs, stu\u00addio cap\u00adture, far-field or set noise con\u00addi\u00adtions<\/td><td>Medi\u00adum<\/td><\/tr><tr><td>5. Vol\u00adume and accept\u00aded-hour rules<\/td><td>Raw record\u00aded hours, loose rules<\/td><td>Net speech hours under strict accep\u00adtance cri\u00adte\u00adria<\/td><td>Medi\u00adum<\/td><\/tr><tr><td>6. Tran\u00adscrip\u00adtion depth<\/td><td>Ortho\u00adgraph\u00adic text, no time\u00adstamps<\/td><td>Ver\u00adba\u00adtim with tags, word time\u00adstamps and diariza\u00adtion<\/td><td>High<\/td><\/tr><tr><td>7. QA depth<\/td><td>Sam\u00adple-based audit<\/td><td>100% sec\u00adond pass, dou\u00adble tran\u00adscrip\u00adtion, WER audits<\/td><td>Medi\u00adum<\/td><\/tr><tr><td>8. Con\u00adsent, meta\u00adda\u00adta and com\u00adpli\u00adance<\/td><td>Sin\u00adgle-pur\u00adpose con\u00adsent, basic meta\u00adda\u00adta<\/td><td>Per-pur\u00adpose con\u00adsent, rich meta\u00adda\u00adta, EU data han\u00addling<\/td><td>Medi\u00adum<\/td><\/tr><tr><td>9. Licens\u00ading and usage rights<\/td><td>Non-exclu\u00adsive license<\/td><td>Exclu\u00adsive rights or full own\u00ader\u00adship<\/td><td>High<\/td><\/tr><tr><td>10. Turn\u00adaround<\/td><td>Flex\u00adi\u00adble sched\u00adule<\/td><td>Rush deliv\u00adery with par\u00adal\u00adlel teams<\/td><td>Low<\/td><\/tr><\/tbody><\/table><figcaption class=\"wp-element-caption\">Speech data cost dri\u00advers and the typ\u00adi\u00adcal swing each one caus\u00ades. Swing is our qual\u00adi\u00adta\u00adtive read, not a mea\u00adsured fig\u00adure.<\/figcaption><\/figure>\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"880\" src=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-cost-drivers.webp\" alt=\"Ten speech data cost drivers rated by relative cost impact: language, speakers, style, channel, volume, transcription, QA, consent, licensing and turnaround\" class=\"wp-image-392\" srcset=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-cost-drivers.webp 1600w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-cost-drivers-300x165.webp 300w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-cost-drivers-1024x563.webp 1024w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-cost-drivers-768x422.webp 768w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-cost-drivers-1536x845.webp 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\"><figcaption class=\"wp-element-caption\">The 10 cost dri\u00advers and how far each can typ\u00adi\u00adcal\u00adly swing a speech data quote (qual\u00adi\u00adta\u00adtive).<\/figcaption><\/figure>\n\n\n<h3 class=\"wp-block-heading\">1. Language and accent rarity<\/h3>\n\n\n<p class=\"wp-block-paragraph\">Recruit\u00ading is the main bot\u00adtle\u00adneck, since speak\u00aders of a wide\u00adly spo\u00adken lan\u00adguage are easy to find. A low-resource lan\u00adguage or a spe\u00adcif\u00adic region\u00adal accent needs local recruiters, native tran\u00adscribers and extra screen\u00ading to con\u00adfirm each speak\u00ader real\u00adly has that accent.<\/p>\n\n\n<p class=\"wp-block-paragraph\"><strong>Spec\u00adi\u00adfy:<\/strong> lan\u00adguage, locale, accent or dialect labels, and how speak\u00ader claims are ver\u00adi\u00adfied. For India-spe\u00adcif\u00adic plan\u00adning, see our <a href=\"https:\/\/www.graveiensai.com\/blog\/indic-speech-data-collection\/\">Indic speech data col\u00adlec\u00adtion guide<\/a>.<\/p>\n\n\n<h3 class=\"wp-block-heading\">2. Speakers and demographic quotas<\/h3>\n\n\n<p class=\"wp-block-paragraph\">More unique speak\u00aders per hour means more recruit\u00ading, onboard\u00ading and con\u00adsent records per hour. Hard quo\u00adtas also slow the end of a project, because the last few cells, such as speak\u00aders over 60 in one region, take the longest to fill.<\/p>\n\n\n<p class=\"wp-block-paragraph\"><strong>Spec\u00adi\u00adfy:<\/strong> total speak\u00aders, a cap on min\u00adutes per speak\u00ader, quo\u00adta cells, and which quo\u00adtas are hard and which are best effort.<\/p>\n\n\n<h3 class=\"wp-block-heading\">3. Speech style: scripted, spontaneous or conversational<\/h3>\n\n\n<p class=\"wp-block-paragraph\">Script\u00aded speech is the fastest to record and to tran\u00adscribe, because the prompt is a start\u00ading tran\u00adscript. Spon\u00adta\u00adneous speech needs top\u00adic prompts and full tran\u00adscrip\u00adtion. Two-par\u00adty con\u00adver\u00adsa\u00adtion adds sched\u00adul\u00ading, a sep\u00ada\u00adrate chan\u00adnel per speak\u00ader and turn-lev\u00adel tran\u00adscrip\u00adtion with over\u00adlaps.<\/p>\n\n\n<p class=\"wp-block-paragraph\">If you are build\u00ading a phone or app agent, read <a href=\"https:\/\/www.graveiensai.com\/blog\/voice-agent-training-data\/\">what voice agent train\u00ading data needs<\/a> before you choose the mix.<\/p>\n\n\n<h3 class=\"wp-block-heading\">4. Channel and recording environment<\/h3>\n\n\n<p class=\"wp-block-paragraph\">Tele\u00adpho\u00adny audio is nar\u00adrow\u00adband, usu\u00adal\u00adly sam\u00adpled at 8 kHz. Wide\u00adband app audio is often 16 kHz, and TTS voic\u00ades are usu\u00adal\u00adly record\u00aded at 44.1 or 48 kHz in treat\u00aded rooms. Each set\u00adup needs dif\u00adfer\u00adent devices, rooms and checks.<\/p>\n\n\n<p class=\"wp-block-paragraph\">Real envi\u00adron\u00adments such as cars, kitchens or streets mean record\u00ading peo\u00adple where they are, then tag\u00adging each file. Con\u00adtrolled noise con\u00addi\u00adtions and far-field micro\u00adphone arrays add book\u00ading and equip\u00adment time. Nev\u00ader accept upsam\u00adpled nar\u00adrow\u00adband audio labeled as wide\u00adband.<\/p>\n\n\n<h3 class=\"wp-block-heading\">5. Volume and the accepted-hour definition<\/h3>\n\n\n<p class=\"wp-block-paragraph\">Larg\u00ader vol\u00adumes spread fixed costs, such as guide\u00adlines, tool\u00ading and pilot work, over more hours. But the def\u00adi\u00adn\u00adi\u00adtion of an hour often moves the price more than the vol\u00adume does. We cov\u00ader this in the next sec\u00adtion.<\/p>\n\n\n<h3 class=\"wp-block-heading\">6. Transcription depth<\/h3>\n\n\n<p class=\"wp-block-paragraph\">An ortho\u00adgraph\u00adic tran\u00adscript records the words, while ver\u00adba\u00adtim also cap\u00adtures fillers, false starts and cut-off words. Each extra lay\u00ader, such as noise and over\u00adlap tags, seg\u00adment or word time\u00adstamps, speak\u00ader diariza\u00adtion and lan\u00adguage tags for code-switch\u00ading, adds a pass and a guide\u00adline to fol\u00adlow. That is why tran\u00adscrip\u00adtion cost per hour of audio climbs fast.<\/p>\n\n\n<p class=\"wp-block-paragraph\">Tran\u00adscrip\u00adtion appears here only as a cost dri\u00adver. If you need it on its own, our <a href=\"https:\/\/www.graveiensai.com\/transcription\">human audio tran\u00adscrip\u00adtion ser\u00advices<\/a> page lists the ver\u00adba\u00adtim and time-aligned options.<\/p>\n\n\n<h3 class=\"wp-block-heading\">7. QA depth and WER audits<\/h3>\n\n\n<p class=\"wp-block-paragraph\">Sam\u00adple-based audits check a share of each batch. Eval\u00adu\u00ada\u00adtion sets often need 100% sec\u00adond-pass review or dou\u00adble tran\u00adscrip\u00adtion. A WER audit has a review\u00ader re-tran\u00adscribe a sam\u00adple, then mea\u00adsures the word error rate between the deliv\u00adered tran\u00adscript and that ref\u00ader\u00adence.<\/p>\n\n\n<p class=\"wp-block-paragraph\"><strong>Spec\u00adi\u00adfy:<\/strong> the audit sam\u00adple size, the met\u00adric, the pass thresh\u00adold per batch and what hap\u00adpens to batch\u00ades that fail.<\/p>\n\n\n<h3 class=\"wp-block-heading\">8. Consent, metadata and compliance<\/h3>\n\n\n<p class=\"wp-block-paragraph\">Per-pur\u00adpose con\u00adsent, a con\u00adsent ID on every file and rich meta\u00adda\u00adta (age band, gen\u00adder, region, device, envi\u00adron\u00adment) all take time to cap\u00adture and check. That work pro\u00adtects you lat\u00ader, when you need to prove cov\u00ader\u00adage or hon\u00ador a dele\u00adtion request.<\/p>\n\n\n<p class=\"wp-block-paragraph\">If any speak\u00aders are in the EU, the GDPR applies to their record\u00adings. Per the <a href=\"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\" target=\"_blank\" rel=\"noopener\">Euro\u00adpean Com\u00admis\u00adsion<\/a>, bio\u00admet\u00adric data processed sole\u00adly to iden\u00adti\u00adfy a per\u00adson is sen\u00adsi\u00adtive data with extra con\u00addi\u00adtions. Our <a href=\"https:\/\/www.graveiensai.com\/blog\/speaker-consent-form-template\/\">speak\u00ader con\u00adsent form tem\u00adplate<\/a> cov\u00aders the details. This is gen\u00ader\u00adal infor\u00adma\u00adtion, not legal advice.<\/p>\n\n\n<h3 class=\"wp-block-heading\">9. Licensing exclusivity and usage rights<\/h3>\n\n\n<p class=\"wp-block-paragraph\">A non-exclu\u00adsive license is cheap\u00adest, because the ven\u00addor can license the same audio to oth\u00aders. Exclu\u00adsive rights or full own\u00ader\u00adship remove that resale val\u00adue, so the price ris\u00ades. Rights for syn\u00adthet\u00adic voice or voice cloning need sep\u00ada\u00adrate speak\u00ader con\u00adsent, which also affects cost.<\/p>\n\n\n<h3 class=\"wp-block-heading\">10. Turnaround<\/h3>\n\n\n<p class=\"wp-block-paragraph\">Rush time\u00adlines mean par\u00adal\u00adlel recruit\u00ading, more review\u00aders and tighter QA win\u00addows. Turn\u00adaround rarely changes the per-hour price as much as the dri\u00advers above, but it can force trade-offs in quo\u00adta bal\u00adance. Build pilot time into your plan.<\/p>\n\n\n<p class=\"wp-block-paragraph\"><strong>In short:<\/strong> four dri\u00advers do most of the work. Pin down lan\u00adguage, speech style, tran\u00adscrip\u00adtion depth and rights first, and the rest of the quote gets much eas\u00adi\u00ader to com\u00adpare.<\/p>\n\n\n<h2 class=\"wp-block-heading\" id=\"what-is-the-difference-between-recorded-and-accepted-hours\">What is the difference between recorded and accepted hours?<\/h2>\n\n\n<p class=\"wp-block-paragraph\">Record\u00aded hours count all cap\u00adtured audio, includ\u00ading silence, retakes and files that fail checks. Accept\u00aded hours count only audio that meets your writ\u00adten accep\u00adtance cri\u00adte\u00adria, such as sam\u00adple rate, noise lev\u00adel, speech con\u00adtent, tran\u00adscript accu\u00adra\u00adcy and com\u00adplete meta\u00adda\u00adta. Pric\u00ading per accept\u00aded hour moves the risk of rejects to the ven\u00addor. The rate looks high\u00ader, but it buys usable data.<\/p>\n\n\n<p class=\"wp-block-paragraph\">It helps to sep\u00ada\u00adrate four kinds of \u201chour\u201d in every con\u00adtract:<\/p>\n\n\n<ol class=\"wp-block-list\"><li><strong>Ses\u00adsion hours.<\/strong> Time speak\u00aders spend in a record\u00ading ses\u00adsion, includ\u00ading set\u00adup, instruc\u00adtions and breaks.<\/li><li><strong>Record\u00aded hours.<\/strong> Total audio cap\u00adtured, includ\u00ading silence and retakes.<\/li><li><strong>Net speech hours.<\/strong> Audio left after trim\u00adming long silences and remov\u00ading retakes.<\/li><li><strong>Accept\u00aded hours.<\/strong> Net speech that pass\u00ades every accep\u00adtance check and is approved by you.<\/li><\/ol>\n\n\n<p class=\"wp-block-paragraph\">The gap between raw and usable hours is real, even out\u00adside cus\u00adtom col\u00adlec\u00adtion. <a href=\"https:\/\/arxiv.org\/abs\/2111.09344\" target=\"_blank\" rel=\"noopener\">Peo\u00adple\u2019s Speech<\/a> start\u00aded with 52,500 hours of audio with tran\u00adscripts. After forced align\u00adment and fil\u00adter\u00ading out mis\u00admatched tran\u00adscripts, 31,400 hours were aligned, about 60% of the source.<\/p>\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"700\" src=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-recorded-vs-accepted-hours.webp\" alt=\"Bar chart from the People's Speech paper: 52,500 source hours produced 31,400 aligned hours, about 60 percent usable\" class=\"wp-image-393\" srcset=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-recorded-vs-accepted-hours.webp 1600w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-recorded-vs-accepted-hours-300x131.webp 300w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-recorded-vs-accepted-hours-1024x448.webp 1024w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-recorded-vs-accepted-hours-768x336.webp 768w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-recorded-vs-accepted-hours-1536x672.webp 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\"><figcaption class=\"wp-element-caption\">Raw hours shrink after qual\u00adi\u00adty fil\u00adter\u00ading: 52,500 source hours became 31,400 aligned hours in Peo\u00adple\u2019s Speech (Galvez et al., 2021).<\/figcaption><\/figure>\n\n\n<p class=\"wp-block-paragraph\">Your yield will dif\u00adfer, but the les\u00adson holds. Write accep\u00adtance cri\u00adte\u00adria before record\u00ading starts, so both sides price the same thing:<\/p>\n\n\n<ul class=\"wp-block-list\"><li>Audio for\u00admat: sam\u00adple rate, bit depth, chan\u00adnel count, file type, no clip\u00adping.<\/li><li>Speech con\u00adtent: min\u00adi\u00admum speech ratio and silence trim\u00adming rules.<\/li><li>Envi\u00adron\u00adment: noise lev\u00adel or envi\u00adron\u00adment tag match\u00ades the spec.<\/li><li>Tran\u00adscript accu\u00adra\u00adcy: max\u00adi\u00admum WER on an audit\u00aded sam\u00adple.<\/li><li>Meta\u00adda\u00adta: every required field filled, plus a valid con\u00adsent ID.<\/li><li>Quo\u00adtas: speak\u00ader caps and quo\u00adta cells respect\u00aded per batch.<\/li><li>Rework: rejec\u00adtion rea\u00adsons, rework rules and dead\u00adlines.<\/li><\/ul>\n\n\n<p class=\"wp-block-paragraph\"><strong>In short:<\/strong> a cheap record\u00aded hour can be an expen\u00adsive accept\u00aded hour. Com\u00adpare speech data col\u00adlec\u00adtion cost per hour on accept\u00aded hours, not record\u00aded ones.<\/p>\n\n\n<h2 class=\"wp-block-heading\" id=\"how-is-a-speech-data-quote-built\">How is a speech data quote built?<\/h2>\n\n\n<p class=\"wp-block-paragraph\">A reli\u00adable quote fol\u00adlows five steps: a writ\u00adten spec, a free sam\u00adple against that spec, a small paid pilot, scaled batch deliv\u00adery and approval of each batch. The pilot is where the per-hour price becomes evi\u00addence-based, because it mea\u00adsures real reject rates, tran\u00adscrip\u00adtion effort and how fast quo\u00adta cells fill.<\/p>\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"720\" src=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-quote-process.webp\" alt=\"Five-step speech data quote process: spec, sample, pilot, scale and approve, with accepted hours growing at each step\" class=\"wp-image-394\" srcset=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-quote-process.webp 1600w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-quote-process-300x135.webp 300w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-quote-process-1024x461.webp 1024w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-quote-process-768x346.webp 768w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-quote-process-1536x691.webp 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\"><figcaption class=\"wp-element-caption\">How a reli\u00adable speech data quote is built, from writ\u00adten spec to approved hours.<\/figcaption><\/figure>\n\n\n<ol class=\"wp-block-list\"><li><strong>Spec.<\/strong> Your team and the ven\u00addor\u2019s project lead agree lan\u00adguages, speak\u00aders, style, audio for\u00admat, tran\u00adscrip\u00adtion guide\u00adline, accep\u00adtance cri\u00adte\u00adria and license. Out\u00adput: a signed spec.<\/li><li><strong>Sam\u00adple.<\/strong> The ven\u00addor records and tran\u00adscribes a short sam\u00adple to that spec. You check audio, tran\u00adscripts and meta\u00adda\u00adta against your own cri\u00adte\u00adria.<\/li><li><strong>Pilot.<\/strong> A small paid batch runs end to end. You mea\u00adsure reject rate, rework time and quo\u00adta fill, then con\u00adfirm or adjust the price.<\/li><li><strong>Scale.<\/strong> Batch\u00ades ship with quo\u00adta reports and QA reports. Native-speak\u00ader review\u00aders audit each batch before deliv\u00adery.<\/li><li><strong>Approve.<\/strong> You accept or reject each batch against the spec. In pay-on-approval mod\u00adels, only approved files are invoiced.<\/li><\/ol>\n\n\n<p class=\"wp-block-paragraph\"><strong>In short:<\/strong> do not sign a large order off a rate card. Let a pilot set the price.<\/p>\n\n\n<h2 class=\"wp-block-heading\" id=\"is-off-the-shelf-speech-data-cheaper-than-custom-collection\">Is off-the-shelf speech data cheaper than custom collection?<\/h2>\n\n\n<p class=\"wp-block-paragraph\">Off-the-shelf speech data is usu\u00adal\u00adly cheap\u00ader and faster up front, because record\u00ading costs are shared across buy\u00aders. Cus\u00adtom col\u00adlec\u00adtion costs more per hour but match\u00ades your accents, chan\u00adnel, domain and rights. The cheap\u00ader option is the one that fits. A free cor\u00adpus that does not match pro\u00adduc\u00adtion audio, or for\u00adbids com\u00admer\u00adcial use, can cost more in the end.<\/p>\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>Fac\u00adtor<\/th><th>Off-the-shelf or open data<\/th><th>Cus\u00adtom col\u00adlec\u00adtion<\/th><\/tr><\/thead><tbody><tr><td>Upfront cost<\/td><td>Low, some\u00adtimes free<\/td><td>High\u00ader, priced to your spec<\/td><\/tr><tr><td>Time to first data<\/td><td>Fast, often imme\u00addi\u00adate<\/td><td>Longer, after a sam\u00adple and pilot<\/td><\/tr><tr><td>Chan\u00adnel and domain fit<\/td><td>What\u00adev\u00ader the cor\u00adpus cap\u00adtured<\/td><td>Matched to your prod\u00aduct<\/td><\/tr><tr><td>Speak\u00ader meta\u00adda\u00adta<\/td><td>Often lim\u00adit\u00aded<\/td><td>Fields you define<\/td><\/tr><tr><td>Com\u00admer\u00adcial rights<\/td><td>Varies: open, research-only or licensed<\/td><td>Writ\u00adten into your con\u00adtract<\/td><\/tr><tr><td>Exclu\u00adsiv\u00adi\u00adty<\/td><td>None, oth\u00aders use the same data<\/td><td>Avail\u00adable at a pre\u00admi\u00adum<\/td><\/tr><\/tbody><\/table><figcaption class=\"wp-element-caption\">Off-the-shelf vs cus\u00adtom speech data at a glance.<\/figcaption><\/figure>\n\n\n<p class=\"wp-block-paragraph\">Open data can be a strong start. <a href=\"https:\/\/arxiv.org\/abs\/2111.09344\" target=\"_blank\" rel=\"noopener\">Peo\u00adple\u2019s Speech<\/a> offers about 30,000 hours of most\u00adly Eng\u00adlish speech under CC-BY-SA, with a CC-BY sub\u00adset, and its paper describes Mozil\u00adla Com\u00admon Voice as CC0-licensed. Both suit pre\u00adtrain\u00ading and research, but check whether their accents, chan\u00adnels and domains match your users.<\/p>\n\n\n<p class=\"wp-block-paragraph\">Licens\u00ading his\u00adto\u00adry mat\u00adters too. The same paper recounts that Baidu col\u00adlect\u00aded 12,000 hours of read speech from 9,600 speak\u00aders through crowd\u00adsourc\u00ading. A legal review lat\u00ader found the speak\u00ader agree\u00adments did not per\u00admit pub\u00adlic release or unre\u00adstrict\u00aded com\u00admer\u00adcial reuse, so the data was nev\u00ader released.<\/p>\n\n\n<p class=\"wp-block-paragraph\">Most teams end up with a hybrid that uses open or licensed data for breadth and cus\u00adtom data for their own domain and chan\u00adnel. They also keep a cus\u00adtom held-out test set with speak\u00aders who nev\u00ader appear in train\u00ading.<\/p>\n\n\n<p class=\"wp-block-paragraph\"><strong>In short:<\/strong> buy breadth, col\u00adlect fit, and always own your test set.<\/p>\n\n\n<h2 class=\"wp-block-heading\" id=\"which-speech-data-licensing-model-fits-your-budget\">Which speech data licensing model fits your budget?<\/h2>\n\n\n<p class=\"wp-block-paragraph\">Pick the nar\u00adrow\u00adest speech data licens\u00ading mod\u00adel that cov\u00aders your real use. Open licens\u00ades cost noth\u00ading but car\u00adry attri\u00adbu\u00adtion or share-alike terms. Research-only licens\u00ades are cheap but block prod\u00aducts. Non-exclu\u00adsive com\u00admer\u00adcial licens\u00ades suit most teams. Exclu\u00adsive licens\u00ades and full own\u00ader\u00adship cost the most, so reserve them for data that gives you a last\u00ading edge.<\/p>\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>Licens\u00ading mod\u00adel<\/th><th>What you get<\/th><th>Cost effect<\/th><\/tr><\/thead><tbody><tr><td>Open license (CC0, CC-BY, CC-BY-SA)<\/td><td>Free use under the license terms, such as attri\u00adbu\u00adtion or share-alike<\/td><td>Low\u00adest<\/td><\/tr><tr><td>Research-only license<\/td><td>Non\u00adcom\u00admer\u00adcial research and edu\u00adca\u00adtion use<\/td><td>Low, but not usable in prod\u00aducts<\/td><\/tr><tr><td>Non-exclu\u00adsive com\u00admer\u00adcial license<\/td><td>Use in prod\u00aducts; the ven\u00addor may license the same data to oth\u00aders<\/td><td>Mod\u00ader\u00adate<\/td><\/tr><tr><td>Time- or field-lim\u00adit\u00aded exclu\u00adsive<\/td><td>No resale to oth\u00aders for a set peri\u00adod or field<\/td><td>High\u00ader<\/td><\/tr><tr><td>Full own\u00ader\u00adship or assign\u00adment<\/td><td>You own the dataset out\u00adright<\/td><td>High\u00adest<\/td><\/tr><\/tbody><\/table><figcaption class=\"wp-element-caption\">Com\u00admon speech data licens\u00ading mod\u00adels and their rel\u00ada\u00adtive cost.<\/figcaption><\/figure>\n\n\n<p class=\"wp-block-paragraph\">What\u00adev\u00ader the mod\u00adel, list the rights you need in writ\u00ading:<\/p>\n\n\n<ul class=\"wp-block-list\"><li>Train\u00ading, fine-tun\u00ading and eval\u00adu\u00ada\u00adtion of your mod\u00adels.<\/li><li>TTS or syn\u00adthet\u00adic voice use, backed by sep\u00ada\u00adrate speak\u00ader con\u00adsent.<\/li><li>Use by affil\u00adi\u00adates, con\u00adtrac\u00adtors and cloud proces\u00adsors.<\/li><li>Deriv\u00ada\u00adtive datasets, such as aug\u00adment\u00aded or fil\u00adtered ver\u00adsions.<\/li><li>Term, ter\u00adri\u00adto\u00adry and what hap\u00adpens to data and mod\u00adels if the license ends.<\/li><\/ul>\n\n\n<p class=\"wp-block-paragraph\">This sec\u00adtion is gen\u00ader\u00adal infor\u00adma\u00adtion, not legal advice. Have coun\u00adsel review any license before you sign.<\/p>\n\n\n<h2 class=\"wp-block-heading\" id=\"the-hours-spec-check-how-to-get-comparable-speech-data-quotes\">The HOURS spec check: how to get comparable speech data quotes<\/h2>\n\n\n<p class=\"wp-block-paragraph\">We use a sim\u00adple tool to turn the ten dri\u00advers into a spec any ven\u00addor can price: the <strong>HOURS spec check<\/strong>. Each let\u00adter cov\u00aders two dri\u00advers, and if any let\u00adter is blank, the quote you get back is a guess.<\/p>\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"800\" src=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-hours-framework.webp\" alt=\"The HOURS spec check framework for speech data quotes: Hours, Ownership, Users, Recording and Standards, each covering two cost drivers\" class=\"wp-image-395\" srcset=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-hours-framework.webp 1600w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-hours-framework-300x150.webp 300w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-hours-framework-1024x512.webp 1024w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-hours-framework-768x384.webp 768w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/speech-data-hours-framework-1536x768.webp 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\"><figcaption class=\"wp-element-caption\">The HOURS spec check groups the 10 cost dri\u00advers into five let\u00adters you can fill in before ask\u00ading for quotes.<\/figcaption><\/figure>\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>Let\u00adter<\/th><th>Cov\u00aders dri\u00advers<\/th><th>Ques\u00adtions to answer<\/th><\/tr><\/thead><tbody><tr><td>H: Hours<\/td><td>Vol\u00adume and accept\u00aded-hour rules; turn\u00adaround<\/td><td>How many accept\u00aded hours? What counts as accept\u00aded? When do batch\u00ades ship?<\/td><\/tr><tr><td>O: Own\u00ader\u00adship<\/td><td>Licens\u00ading rights; con\u00adsent and meta\u00adda\u00adta<\/td><td>Exclu\u00adsive or not? Which uses? Which con\u00adsent pur\u00adpos\u00ades and meta\u00adda\u00adta fields?<\/td><\/tr><tr><td>U: Users<\/td><td>Lan\u00adguage and accent; speak\u00aders and quo\u00adtas<\/td><td>Which lan\u00adguages and accents? How many speak\u00aders? Which quo\u00adtas are hard?<\/td><\/tr><tr><td>R: Record\u00ading<\/td><td>Speech style; chan\u00adnel and envi\u00adron\u00adment<\/td><td>Script\u00aded, spon\u00adta\u00adneous or con\u00adver\u00adsa\u00adtion\u00adal? Which sam\u00adple rate, devices and set\u00adtings?<\/td><\/tr><tr><td>S: Stan\u00addards<\/td><td>Tran\u00adscrip\u00adtion depth; QA depth<\/td><td>Which tran\u00adscrip\u00adtion guide\u00adline and tags? Which QA met\u00adric and thresh\u00adold?<\/td><\/tr><\/tbody><\/table><figcaption class=\"wp-element-caption\">The HOURS spec check maps each of the 10 cost dri\u00advers to a ques\u00adtion you can answer before request\u00ading quotes.<\/figcaption><\/figure>\n\n\n<h3 class=\"wp-block-heading\">HOURS checklist: spec fields to fill in<\/h3>\n\n\n<p class=\"wp-block-paragraph\">Copy these fields into your request for quote. Exam\u00adple entries show the lev\u00adel of detail that gets a pre\u00adcise answer.<\/p>\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>Spec field<\/th><th>Exam\u00adple entry<\/th><\/tr><\/thead><tbody><tr><td>Use case<\/td><td>ASR fine-tun\u00ading for a cus\u00adtomer sup\u00adport voice agent<\/td><\/tr><tr><td>Lan\u00adguages and locales<\/td><td>US Eng\u00adlish; Ital\u00adian (Italy)<\/td><\/tr><tr><td>Accents or dialects<\/td><td>US South\u00adern, US Mid\u00adwest; North\u00adern and South\u00adern Ital\u00adian<\/td><\/tr><tr><td>Accept\u00aded hours<\/td><td>200 accept\u00aded hours per lan\u00adguage, net of silence<\/td><\/tr><tr><td>Speak\u00aders and caps<\/td><td>At least 500 speak\u00aders per lan\u00adguage, 30 min\u00adutes max\u00adi\u00admum each<\/td><\/tr><tr><td>Quo\u00adtas<\/td><td>Age bands 18\u201330, 31\u201350, 51+; gen\u00adder with\u00adin an agreed tol\u00ader\u00adance<\/td><\/tr><tr><td>Speech style mix<\/td><td>20% script\u00aded, 30% spon\u00adta\u00adneous, 50% two-par\u00adty role-play<\/td><\/tr><tr><td>Chan\u00adnel and for\u00admat<\/td><td>8 kHz tele\u00adpho\u00adny pairs, one speak\u00ader per chan\u00adnel, 16-bit WAV<\/td><\/tr><tr><td>Envi\u00adron\u00adments<\/td><td>Qui\u00adet room, home with TV, car, street<\/td><\/tr><tr><td>Tran\u00adscrip\u00adtion<\/td><td>Ver\u00adba\u00adtim, noise and over\u00adlap tags, seg\u00adment time\u00adstamps, speak\u00ader turns<\/td><\/tr><tr><td>QA and accep\u00adtance<\/td><td>Audit\u00aded sam\u00adple per batch, WER thresh\u00adold agreed after pilot<\/td><\/tr><tr><td>Meta\u00adda\u00adta and con\u00adsent<\/td><td>Speak\u00ader ID, age band, gen\u00adder, region, device, envi\u00adron\u00adment, con\u00adsent ID<\/td><\/tr><tr><td>License<\/td><td>Non-exclu\u00adsive com\u00admer\u00adcial license for train\u00ading and eval\u00adu\u00ada\u00adtion<\/td><\/tr><tr><td>Time\u00adline<\/td><td>Sam\u00adple, pilot, then week\u00adly batch\u00ades<\/td><\/tr><\/tbody><\/table><figcaption class=\"wp-element-caption\">The HOURS check\u00adlist spec fields with illus\u00adtra\u00adtive exam\u00adple entries. Replace every entry with your own require\u00adments.<\/figcaption><\/figure>\n\n\n<p class=\"cta-line wp-block-paragraph\"><strong>Send your spec, get a scoped quote.<\/strong> Fill in the HOURS fields and <a href=\"https:\/\/www.graveiensai.com\/contact-us\">send them to our voice data team<\/a>. We will scope a pilot billed only on approved files. <a href=\"https:\/\/www.graveiensai.com\/contact-us\"><strong>Request the HOURS check\u00adlist<\/strong><\/a> (men\u00adtion \u201cHOURS check\u00adlist\u201d in your mes\u00adsage and we\u2019ll email the one-page PDF).<\/p>\n\n\n<h2 class=\"wp-block-heading\" id=\"common-mistakes-that-inflate-speech-data-costs\">Common mistakes that inflate speech data costs<\/h2>\n\n\n<p class=\"wp-block-paragraph\">Most over\u00adruns in the cost of audio data col\u00adlec\u00adtion come from deci\u00adsions made before the first file is record\u00aded. These are the ones we see most often:<\/p>\n\n\n<ul class=\"wp-block-list\"><li><strong>Com\u00adpar\u00ading dif\u00adfer\u00adent hours.<\/strong> A record\u00aded-hour quote and an accept\u00aded-hour quote look alike on paper but buy dif\u00adfer\u00adent amounts of usable audio.<\/li><li><strong>Writ\u00ading accep\u00adtance rules at deliv\u00adery.<\/strong> If cri\u00adte\u00adria arrive late, reject\u00aded batch\u00ades become dis\u00adputes instead of rework.<\/li><li><strong>Record\u00ading the wrong chan\u00adnel.<\/strong> Stu\u00addio audio for a phone prod\u00aduct, or phone audio for a TTS voice, means pay\u00ading twice.<\/li><li><strong>Over-tran\u00adscrib\u00ading train\u00ading data.<\/strong> Bulk train\u00ading audio may only need ortho\u00adgraph\u00adic text. Save ver\u00adba\u00adtim, tags and dou\u00adble review for the eval\u00adu\u00ada\u00adtion set.<\/li><li><strong>Loose quo\u00adtas.<\/strong> Vague demo\u00adgraph\u00adic tar\u00adgets skew toward the eas\u00adi\u00adest speak\u00aders to recruit, and re-col\u00adlect\u00ading the gaps costs more lat\u00ader.<\/li><li><strong>Buy\u00ading research-only data for a prod\u00aduct.<\/strong> A low price means lit\u00adtle if the license for\u00adbids com\u00admer\u00adcial use.<\/li><li><strong>Pay\u00ading for exclu\u00adsiv\u00adi\u00adty you do not need.<\/strong> Exclu\u00adsive rights make sense for dif\u00adfer\u00aden\u00adti\u00adat\u00ading data, not for gener\u00adic read speech.<\/li><li><strong>Skip\u00adping the pilot.<\/strong> With\u00adout a pilot, nobody knows the real reject rate, so both sides guess and pad the price.<\/li><\/ul>\n\n\n<p class=\"wp-block-paragraph\"><strong>In short:<\/strong> the cheap\u00adest hour is the one you only pay for once.<\/p>\n\n\n<h2 class=\"wp-block-heading\" id=\"how-graveiens-ai-quotes-speech-data-collection\">How Graveiens AI quotes speech data collection<\/h2>\n\n\n<p class=\"wp-block-paragraph\">Graveiens AI does not pub\u00adlish a rate card, because your spec sets your speech data col\u00adlec\u00adtion cost per hour. We quote against your writ\u00adten spec, bro\u00adken into line items you can com\u00adpare with any oth\u00ader ven\u00addor.<\/p>\n\n\n<ul class=\"wp-block-list\"><li><strong>Free 30-minute speech sam\u00adple set.<\/strong> Record\u00aded and tran\u00adscribed to your spec, so you can check qual\u00adi\u00adty before a pilot.<\/li><li><strong>Pilot billed only on approved files.<\/strong> You are invoiced only for deliv\u00ader\u00adables you approve, at the pilot stage and after.<\/li><li><strong>Explic\u00adit-con\u00adsent work\u00adflow.<\/strong> Par\u00adtic\u00adi\u00adpant onboard\u00ading, then record\u00ading man\u00adage\u00adment, then meta\u00adda\u00adta tag\u00adging, then QA, with a con\u00adsent record behind every file.<\/li><li><strong>Lan\u00adguage cov\u00ader\u00adage.<\/strong> 25+ lan\u00adguages, includ\u00ading Indi\u00adan lan\u00adguages and Euro\u00adpean lan\u00adguages such as Ital\u00adian.<\/li><li><strong>Qual\u00adi\u00adty sys\u00adtem.<\/strong> Graveiens AI is ISO 9001 cer\u00adti\u00adfied.<\/li><\/ul>\n\n\n<p class=\"wp-block-paragraph\">Script\u00aded, spon\u00adta\u00adneous and con\u00adver\u00adsa\u00adtion\u00adal col\u00adlec\u00adtion are all avail\u00adable through our <a href=\"https:\/\/www.graveiensai.com\/voice-speech\">con\u00adsent-first voice and speech data ser\u00advices<\/a>, and mul\u00adti-modal pro\u00adgrams run through our <a href=\"https:\/\/www.graveiensai.com\/data-collection\">cus\u00adtom data col\u00adlec\u00adtion pro\u00adgrams<\/a>. For any\u00adthing out\u00adside a stan\u00addard spec, we scope it with you.<\/p>\n\n\n<p class=\"wp-block-paragraph\">We do not promise mod\u00adel accu\u00adra\u00adcy gains, because those depend on your archi\u00adtec\u00adture and eval\u00adu\u00ada\u00adtion. What we com\u00admit to is data that match\u00ades the spec you sign off.<\/p>\n\n\n<h2 class=\"wp-block-heading\" id=\"faq\">Frequently asked questions<\/h2>\n\n<div id=\"rank-math-faq\" class=\"rank-math-block\">\n<div class=\"rank-math-list \">\n<div id=\"faq-question-17950000000190\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">What is a typical speech data collection cost per hour?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>There is no reli\u00adable typ\u00adi\u00adcal fig\u00adure, because the spec dri\u00adves the price. Pub\u00adlic ref\u00ader\u00adence points exist: the Peo\u00adple\u2019s Speech team esti\u00admat\u00aded hand-label\u00ading 100,000 hours at $5 mil\u00adlion in 2021, about $50 per hour for label\u00ading alone. Treat such fig\u00adures as con\u00adtext, and price your own spec with a sam\u00adple and a pilot.<\/p>\n\n<\/div>\n<\/div>\n<div id=\"faq-question-17950000000191\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">What is an accepted hour in speech data collection?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>An accept\u00aded hour is an hour of net speech audio that pass\u00ades every accep\u00adtance check in your spec. Checks cov\u00ader audio for\u00admat, noise lev\u00adel, tran\u00adscript accu\u00adra\u00adcy, meta\u00adda\u00adta com\u00adplete\u00adness and speak\u00ader caps. It excludes silence, retakes and reject\u00aded files. Pric\u00ading per accept\u00aded hour puts the cost of rejects on the ven\u00addor.<\/p>\n\n<\/div>\n<\/div>\n<div id=\"faq-question-17950000000192\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">Is transcription included in speech data collection pricing?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>Some\u00adtimes, and that is a com\u00admon source of con\u00adfu\u00adsion. Some quotes cov\u00ader record\u00ading only, some include machine drafts, and some include human ver\u00adba\u00adtim tran\u00adscrip\u00adtion with time\u00adstamps and tags. Ask for tran\u00adscrip\u00adtion as a sep\u00ada\u00adrate line item, with the guide\u00adline and QA thresh\u00adold named, so you can com\u00adpare quotes fair\u00adly.<\/p>\n\n<\/div>\n<\/div>\n<div id=\"faq-question-17950000000193\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">Why does conversational speech data cost more than read speech?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>Con\u00adver\u00adsa\u00adtion\u00adal data needs two speak\u00aders sched\u00aduled togeth\u00ader, a sep\u00ada\u00adrate chan\u00adnel for each, and sce\u00adnar\u00adios that pro\u00adduce nat\u00adur\u00adal turn-tak\u00ading. Tran\u00adscrip\u00adtion is hard\u00ader too, with over\u00adlaps, fillers and speak\u00ader turns to mark. Read speech starts from a known script, so both record\u00ading and tran\u00adscrip\u00adtion take less effort per hour.<\/p>\n\n<\/div>\n<\/div>\n<div id=\"faq-question-17950000000194\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">Does exclusive licensing increase the cost of speech data?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>Yes. With a non-exclu\u00adsive license, the ven\u00addor can license the same audio to oth\u00ader buy\u00aders, which spreads the cost. Exclu\u00adsive rights or full own\u00ader\u00adship remove that option, so the price ris\u00ades. Reserve exclu\u00adsiv\u00adi\u00adty for data that gives you a last\u00ading advan\u00adtage, such as your own domain or a rare accent.<\/p>\n\n<\/div>\n<\/div>\n<div id=\"faq-question-17950000000195\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">How can I lower speech data collection cost without hurting quality?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>Define accept\u00aded hours and accep\u00adtance cri\u00adte\u00adria up front, match the chan\u00adnel to your prod\u00aduct, and tran\u00adscribe bulk train\u00ading audio at a lighter lev\u00adel than your eval\u00adu\u00ada\u00adtion set. Use open or licensed data for breadth, run a pilot before scal\u00ading, and only pay for exclu\u00adsiv\u00adi\u00adty where it mat\u00adters.<\/p>\n\n<\/div>\n<\/div>\n<div id=\"faq-question-17950000000196\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">Is open speech data free to use commercially?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>Not always. Licens\u00ades dif\u00adfer. CC0 and CC-BY gen\u00ader\u00adal\u00adly allow com\u00admer\u00adcial use, CC-BY-SA adds share-alike terms, and research-only licens\u00ades block com\u00admer\u00adcial use. LDC, for exam\u00adple, lim\u00adits non\u00admem\u00adber licensees to non\u00adcom\u00admer\u00adcial research. Read the license for each cor\u00adpus and have coun\u00adsel con\u00adfirm your intend\u00aded use.<\/p>\n\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n\n\n<div class=\"wp-block-group cta-box\"><div class=\"wp-block-group__inner-container is-layout-flow wp-block-group-is-layout-flow\">\n<h2 class=\"wp-block-heading\" id=\"get-a-free-30-minute-speech-sample-set\">Get a free 30-minute speech sample set<\/h2>\n\n\n<p class=\"wp-block-paragraph\">Tell us your lan\u00adguage, accent, speech style and chan\u00adnel. We will record and tran\u00adscribe a free 30-minute sam\u00adple set to your spec, so you can check it against your own accep\u00adtance cri\u00adte\u00adria. If you go ahead, the pilot is billed only on approved files.<\/p>\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/www.graveiensai.com\/contact-us\"><strong>Request your free speech sam\u00adple set<\/strong><\/a><\/p>\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/www.graveiensai.com\/contact-us\"><strong>Request the HOURS check\u00adlist<\/strong><\/a> (men\u00adtion \u201cHOURS check\u00adlist\u201d in your mes\u00adsage and we\u2019ll email the one-page PDF).<\/p>\n\n<\/div><\/div>\n\n\n<h2 class=\"wp-block-heading\" id=\"sources\">Sources<\/h2>\n\n\n<ol class=\"wp-block-list\"><li>Galvez, D. et al. The Peo\u00adple\u2019s Speech: A Large-Scale Diverse Eng\u00adlish Speech Recog\u00adni\u00adtion Dataset for Com\u00admer\u00adcial Usage. NeurIPS 2021 Datasets and Bench\u00admarks Track, arXiv:2111.09344. <a href=\"https:\/\/arxiv.org\/abs\/2111.09344\" target=\"_blank\" rel=\"noopener\">https:\/\/arxiv.org\/abs\/2111.09344<\/a><\/li><li>Diar\u00adra, Y., Coulibaly, N. S. and Lev\u00aden\u00adthal, M. Cost Analy\u00adsis of Human-cor\u00adrect\u00aded Tran\u00adscrip\u00adtion for Pre\u00addom\u00adi\u00adnate\u00adly Oral Lan\u00adguages. arXiv:2510.12781, Octo\u00adber 2025. <a href=\"https:\/\/arxiv.org\/abs\/2510.12781\" target=\"_blank\" rel=\"noopener\">https:\/\/arxiv.org\/abs\/2510.12781<\/a><\/li><li>Lin\u00adguis\u00adtic Data Con\u00adsor\u00adtium, Uni\u00adver\u00adsi\u00adty of Penn\u00adsyl\u00adva\u00adnia. Licens\u00ading. <a href=\"https:\/\/www.ldc.upenn.edu\/data-management\/using\/licensing\" target=\"_blank\" rel=\"noopener\">https:\/\/www.ldc.upenn.edu\/data-management\/using\/licensing<\/a><\/li><li>Euro\u00adpean Com\u00admis\u00adsion. What per\u00adson\u00adal data is con\u00adsid\u00adered sen\u00adsi\u00adtive? <a href=\"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\" target=\"_blank\" rel=\"noopener\">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<\/a><\/li><\/ol>\n\n\n<p class=\"author-bio wp-block-paragraph\"><strong>About the author: Graveiens AI Edi\u00adto\u00adr\u00adi\u00adal Team.<\/strong> The Graveiens AI edi\u00adto\u00adr\u00adi\u00adal team writes about speech data, AI train\u00ading data and data com\u00adpli\u00adance for AI builders in the US and EU. The team works along\u00adside our voice col\u00adlec\u00adtion, tran\u00adscrip\u00adtion and QA leads, so each guide reflects how real projects are scoped and deliv\u00adered.<\/p>\n\n","protected":false},"excerpt":{"rendered":"<p>Speech data col\u00adlec\u00adtion cost per hour explained: 10 price dri\u00advers, record\u00aded vs accept\u00aded hours, licens\u00ading and a spec check\u00adlist. Get a free speech sam\u00adple.<\/p>\n","protected":false},"author":1,"featured_media":391,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"wp_typography_post_enhancements_disabled":false,"footnotes":""},"categories":[10],"tags":[64,63,14,23],"class_list":["post-396","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-voice-speech-data","tag-asr-training-data","tag-speech-data-pricing","tag-voice-ai","tag-voice-data-collection"],"_links":{"self":[{"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/posts\/396","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/comments?post=396"}],"version-history":[{"count":1,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/posts\/396\/revisions"}],"predecessor-version":[{"id":397,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/posts\/396\/revisions\/397"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/media\/391"}],"wp:attachment":[{"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/media?parent=396"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/categories?post=396"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/tags?post=396"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}