{"id":325,"date":"2026-11-05T18:30:00","date_gmt":"2026-11-05T13:00:00","guid":{"rendered":"https:\/\/www.graveiensai.com\/blog\/?p=325"},"modified":"2026-11-02T07:52:09","modified_gmt":"2026-11-02T02:22:09","slug":"voice-agent-training-data","status":"publish","type":"post","link":"https:\/\/www.graveiensai.com\/blog\/voice-agent-training-data\/","title":{"rendered":"Voice Agent Training Data: What Conversational Speech Datasets Need"},"content":{"rendered":"\n<p class=\"last-updated wp-block-paragraph\">Last updat\u00aded: Novem\u00adber 2026<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Key takeaways<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Voice agent train\u00ading data is two-sided, con\u00adver\u00adsa\u00adtion\u00adal audio with tran\u00adscripts and labels. It must match the chan\u00adnel, accents, noise and turn-tak\u00ading your agent will face live.<\/li>\n\n\n<li>Each part of the stack needs dif\u00adfer\u00adent data: ASR needs tran\u00adscribed audio, the dia\u00adlog lay\u00ader needs labeled con\u00adver\u00adsa\u00adtions, and TTS needs clean, con\u00adsent\u00aded voice record\u00adings.<\/li>\n\n\n<li>Phone agents need 8 kHz tele\u00adpho\u00adny audio in the mix. Record each speak\u00ader on a sep\u00ada\u00adrate chan\u00adnel, and nev\u00ader upsam\u00adple nar\u00adrow\u00adband audio and call it wide\u00adband.<\/li>\n\n\n<li>Word error rate (WER) is the stan\u00addard ASR met\u00adric, but it depends on clear tran\u00adscrip\u00adtion and nor\u00admal\u00adiza\u00adtion rules.<\/li>\n\n\n<li>Voice can be bio\u00admet\u00adric data. Col\u00adlect with explic\u00adit, doc\u00adu\u00adment\u00aded con\u00adsent under GDPR, BIPA-style laws and the CCPA.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Many voice agents sound great in a demo and stum\u00adble on the first real call. The caller is on a noisy phone line, speaks with a region\u00adal accent, and talks over the agent mid-sen\u00adtence. The mod\u00adel nev\u00ader heard any\u00adthing like that in train\u00ading.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Voice agent train\u00ading data clos\u00ades that gap, if it is built for real con\u00adver\u00adsa\u00adtions. This guide cov\u00aders what ASR, dia\u00adlog and TTS mod\u00adels each need, how to plan cov\u00ader\u00adage, and how to col\u00adlect, tran\u00adscribe and check speech data with con\u00adsent in place.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Want to hear the dif\u00adfer\u00adence first? <a href=\"https:\/\/www.graveiensai.com\/contact-us\">Request our free 30-minute speech sam\u00adple set<\/a> in the lan\u00adguage and set\u00adting you care about.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">What is voice agent training data?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Voice agent train\u00ading data is record\u00aded speech, usu\u00adal\u00adly two-sided con\u00adver\u00adsa\u00adtions, paired with time-aligned tran\u00adscripts, speak\u00ader labels and task labels. It teach\u00ades a phone or app agent to hear callers, fol\u00adlow the dia\u00adlog and respond. Good data mir\u00adrors live use: real chan\u00adnels, accents, back\u00adground noise, inter\u00adrup\u00adtions and the tasks callers actu\u00adal\u00adly bring.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">In short: gen\u00ader\u00adal speech data teach\u00ades a mod\u00adel to tran\u00adscribe. Voice agent data teach\u00ades it to hold a con\u00adver\u00adsa\u00adtion.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The dif\u00adfer\u00adence shows up in the details. A read-speech cor\u00adpus has one speak\u00ader, clean audio and com\u00adplete sen\u00adtences. A real sup\u00adport call has two speak\u00aders, over\u00adlaps, fillers like \u201cum,\u201d half-fin\u00adished thoughts and long paus\u00ades. Your data needs both, in pro\u00adpor\u00adtions that match your use case.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">What data does each part of a voice agent need?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Most voice agents chain three mod\u00adels. Speech recog\u00adni\u00adtion (ASR) turns caller audio into text. A lan\u00adguage mod\u00adel or dia\u00adlog man\u00adag\u00ader decides what to say. Text-to-speech (TTS) speaks the reply. Each lay\u00ader has its own data needs, and a gap in one shows up as a fail\u00adure in the whole call.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"780\" src=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-agent-pipeline-data-needs.webp\" alt=\"Voice agent pipeline from caller audio to ASR, LLM dialog and TTS reply, with the training data each stage needs\" class=\"wp-image-320\" srcset=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-agent-pipeline-data-needs.webp 1600w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-agent-pipeline-data-needs-300x146.webp 300w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-agent-pipeline-data-needs-1024x499.webp 1024w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-agent-pipeline-data-needs-768x374.webp 768w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-agent-pipeline-data-needs-1536x749.webp 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\"><figcaption class=\"wp-element-caption\">Caller audio flows through ASR, the LLM or dia\u00adlog lay\u00ader and TTS. Each stage needs its own data: caller audio and tran\u00adscripts, dia\u00adlog trees, and stu\u00addio voic\u00ades.<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-table\"><table><thead>\n<tr><th>Com\u00adpo\u00adnent<\/th><th>Data it needs<\/th><th>Com\u00admon gap<\/th><\/tr>\n<\/thead>\n<tbody>\n<tr><td>ASR<\/td><td>Audio with ver\u00adba\u00adtim, time-aligned tran\u00adscripts<\/td><td>Too lit\u00adtle phone-line, accent\u00aded or noisy speech<\/td><\/tr>\n<tr><td>Turn-tak\u00ading and end\u00adpoint\u00ading<\/td><td>Two-chan\u00adnel audio with over\u00adlap and pause tim\u00ading<\/td><td>Sin\u00adgle-chan\u00adnel mix\u00ades that hide who spoke when<\/td><\/tr>\n<tr><td>LLM or dia\u00adlog lay\u00ader<\/td><td>Mul\u00adti-turn tran\u00adscripts with intents, enti\u00adties and out\u00adcomes<\/td><td>Only \u201chap\u00adpy path\u201d dialogs, no cor\u00adrec\u00adtions or con\u00adfu\u00adsion<\/td><\/tr>\n<tr><td>TTS<\/td><td>Clean stu\u00addio record\u00adings with exact scripts<\/td><td>Miss\u00ading voice-tal\u00adent con\u00adsent for syn\u00adthet\u00adic voice use<\/td><\/tr>\n<tr><td>Eval\u00adu\u00ada\u00adtion<\/td><td>Held-out calls that match pro\u00adduc\u00adtion<\/td><td>Test\u00ading on the same speak\u00aders used in train\u00ading<\/td><\/tr>\n<\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">For speech-to-speech mod\u00adels that skip the text lay\u00ader, the same rules apply. They still need nat\u00adur\u00adal two-sided audio, with each speak\u00ader on a sep\u00ada\u00adrate chan\u00adnel.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">How is voice agent training data collected?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A reli\u00adable col\u00adlec\u00adtion runs in six steps: con\u00adsent, script design, record\u00ading, tran\u00adscrip\u00adtion, QA and deliv\u00adery. Con\u00adsent comes first because it shapes every\u00adthing else, from what you can record to how long you can keep it. Skip\u00adping a step usu\u00adal\u00adly means re-record\u00ading lat\u00ader, which costs more than doing it right once.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"640\" src=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-data-collection-workflow.webp\" alt=\"Voice data collection workflow: consent, script design, recording, transcription, QA and delivery\" class=\"wp-image-321\" srcset=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-data-collection-workflow.webp 1600w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-data-collection-workflow-300x120.webp 300w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-data-collection-workflow-1024x410.webp 1024w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-data-collection-workflow-768x307.webp 768w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-data-collection-workflow-1536x614.webp 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\"><figcaption class=\"wp-element-caption\">Six-step voice data col\u00adlec\u00adtion work\u00adflow, from con\u00adsent to deliv\u00adery.<\/figcaption><\/figure>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>Con\u00adsent.<\/strong> Explain the pur\u00adpose, the uses, reten\u00adtion and with\u00addraw\u00adal. Cap\u00adture a signed record for each speak\u00ader. Our <a href=\"https:\/\/www.graveiensai.com\/blog\/speaker-consent-form-template\/\">speak\u00ader con\u00adsent form tem\u00adplate<\/a> lists the fields to include.<\/li>\n\n\n<li><strong>Script design.<\/strong> Write sce\u00adnar\u00adios, prompts or role-play briefs from your real call rea\u00adsons. Include names, num\u00adbers, address\u00ades and prod\u00aduct terms.<\/li>\n\n\n<li><strong>Record\u00ading.<\/strong> Cap\u00adture on the tar\u00adget chan\u00adnels and devices, with each speak\u00ader on a sep\u00ada\u00adrate track. Log device, envi\u00adron\u00adment and sam\u00adple rate as meta\u00adda\u00adta.<\/li>\n\n\n<li><strong>Tran\u00adscrip\u00adtion.<\/strong> Tran\u00adscribe ver\u00adba\u00adtim to a writ\u00adten style guide, with speak\u00ader turns, time\u00adstamps and tags for noise and over\u00adlap.<\/li>\n\n\n<li><strong>QA.<\/strong> Check audio qual\u00adi\u00adty, tran\u00adscript accu\u00adra\u00adcy, label con\u00adsis\u00adten\u00adcy and meta\u00adda\u00adta com\u00adplete\u00adness on a defined sam\u00adple.<\/li>\n\n\n<li><strong>Deliv\u00adery.<\/strong> Ship audio, tran\u00adscripts, meta\u00adda\u00adta and con\u00adsent ref\u00ader\u00adences in an agreed for\u00admat, with a short QA report.<\/li>\n<\/ol>\n\n\n\n<h2 class=\"wp-block-heading\">The Six-Axis Voice Coverage Matrix<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Cov\u00ader\u00adage is where most voice datasets fall short. We plan col\u00adlec\u00adtions with a sim\u00adple tool: the <strong>Six-Axis Voice Cov\u00ader\u00adage Matrix<\/strong>. List your tar\u00adget val\u00adues on each axis, then fill a grid to check that every com\u00adbi\u00adna\u00adtion that mat\u00adters has enough exam\u00adples. The six axes are accents, chan\u00adnels, noise, intents, dis\u00adflu\u00aden\u00adcies and lan\u00adguages.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"900\" src=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-data-coverage-matrix.webp\" alt=\"Voice data coverage matrix across accents, channels (8 kHz phone vs wideband), noise, intents, disfluencies and languages\" class=\"wp-image-322\" srcset=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-data-coverage-matrix.webp 1600w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-data-coverage-matrix-300x169.webp 300w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-data-coverage-matrix-1024x576.webp 1024w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-data-coverage-matrix-768x432.webp 768w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/voice-data-coverage-matrix-1536x864.webp 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\"><figcaption class=\"wp-element-caption\">The Six-Axis Voice Cov\u00ader\u00adage Matrix: accents, chan\u00adnels, noise, intents, dis\u00adflu\u00aden\u00adcies and lan\u00adguages.<\/figcaption><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">1. Accents<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">List the region\u00adal and sec\u00adond-lan\u00adguage accents of your real callers, not just the \u201cstan\u00addard\u201d one. Spread record\u00adings across many speak\u00aders rather than a few long ses\u00adsions. More speak\u00aders usu\u00adal\u00adly means wider cov\u00ader\u00adage of voic\u00ades and speak\u00ading styles.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2. Channels: 8 kHz telephony vs wideband<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Clas\u00adsic phone calls car\u00adry nar\u00adrow\u00adband audio, sam\u00adpled at 8 kHz. App and web calls often use wide\u00adband codecs, such as ITU\u2011T G.722, which car\u00adries audio up to about 7 kHz. As of Novem\u00adber 2026, NIST\u2019s OpenASR21 eval\u00adu\u00ada\u00adtion plan still describes con\u00adver\u00adsa\u00adtion\u00adal tele\u00adphone speech as two chan\u00adnels, dis\u00adtrib\u00aduted sep\u00ada\u00adrate\u00adly, at 8 kHz or high\u00ader sam\u00adple rates.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Three rules help. Record each speak\u00ader on a sep\u00ada\u00adrate chan\u00adnel. Keep the native sam\u00adple rate in meta\u00adda\u00adta. And nev\u00ader upsam\u00adple 8 kHz audio and label it wide\u00adband, because the miss\u00ading fre\u00adquen\u00adcies do not come back.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3. Noise<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Record in the places callers real\u00adly are: cars, kitchens, streets, open offices and speak\u00ader\u00adphones. Added syn\u00adthet\u00adic noise helps, but it rarely cap\u00adtures echo, crosstalk and phone-line arti\u00adfacts. Tag the envi\u00adron\u00adment for every file.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4. Intents<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Map scripts to your top call rea\u00adsons and the enti\u00adties inside them. Include hard cas\u00ades: spelled names, account num\u00adbers, dates, address\u00ades and cor\u00adrec\u00adtions such as \u201cno, I said fif\u00adteen.\u201d<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5. Disfluencies and turn-taking<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Real speech is messy. Cap\u00adture fillers, false starts, repairs, backchan\u00adnels like \u201cmm-hmm,\u201d long paus\u00ades and barge-in, when a caller talks over the agent. Sep\u00ada\u00adrate chan\u00adnels let you label over\u00adlap tim\u00ading, which turn-tak\u00ading and end\u00adpoint\u00ading mod\u00adels need.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">6. Languages<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">List each lan\u00adguage and locale, plus code-switch\u00ading where callers mix lan\u00adguages. A Span\u00adish line in Texas and one in Madrid need dif\u00adfer\u00adent speak\u00aders. The same applies to Ital\u00adian, Hin\u00addi or French across regions.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Need cov\u00ader\u00adage across sev\u00ader\u00adal lan\u00adguages or accents? <a href=\"https:\/\/www.graveiensai.com\/contact-us\">Scope a pilot billed only on approved files<\/a>, or start with the free 30-minute speech sam\u00adple set.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Should you use scripted, role-play or spontaneous speech?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Use all three, in pro\u00adpor\u00adtions that match your goal. Script\u00aded speech gives con\u00adtrol and exact tran\u00adscripts. Role-play adds nat\u00adur\u00adal turn-tak\u00ading around real tasks. Spon\u00adta\u00adneous con\u00adver\u00adsa\u00adtion gives the most real\u00adism but the least con\u00adtrol. Most voice agent projects lean on role-play, then add script\u00aded prompts for hard enti\u00adties and spon\u00adta\u00adneous speech for real\u00adism.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"860\" src=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/scripted-vs-spontaneous-speech-data.webp\" alt=\"Scripted, role-play and spontaneous speech data compared on a slider from control to realism\" class=\"wp-image-323\" srcset=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/scripted-vs-spontaneous-speech-data.webp 1600w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/scripted-vs-spontaneous-speech-data-300x161.webp 300w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/scripted-vs-spontaneous-speech-data-1024x550.webp 1024w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/scripted-vs-spontaneous-speech-data-768x413.webp 768w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/11\/scripted-vs-spontaneous-speech-data-1536x826.webp 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\"><figcaption class=\"wp-element-caption\">Script\u00aded, role-play and spon\u00adta\u00adneous speech sit on a slid\u00ader from con\u00adtrol to real\u00adism.<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-table\"><table><thead>\n<tr><th>Type<\/th><th>Best for<\/th><th>Watch out for<\/th><\/tr>\n<\/thead>\n<tbody>\n<tr><td>Script\u00aded<\/td><td>Enti\u00adties, wake words, TTS, exact cov\u00ader\u00adage<\/td><td>Read\u00ading voice, few dis\u00adflu\u00aden\u00adcies<\/td><\/tr>\n<tr><td>Role-play<\/td><td>Task dialogs, turn-tak\u00ading, barge-in<\/td><td>Actors drift\u00ading from real caller behav\u00adior<\/td><\/tr>\n<tr><td>Spon\u00adta\u00adneous<\/td><td>Nat\u00adur\u00adal speech, fillers, over\u00adlaps, eval\u00adu\u00ada\u00adtion<\/td><td>Uneven cov\u00ader\u00adage, heav\u00adier tran\u00adscrip\u00adtion work<\/td><\/tr>\n<\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">What\u00adev\u00ader the mix, keep a held-out test set of speak\u00aders who nev\u00ader appear in train\u00ading data.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">How do you measure voice data quality?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Start with clear tran\u00adscrip\u00adtion rules, then mea\u00adsure. Word error rate (WER) is the stan\u00addard ASR met\u00adric. NIST defines it as dele\u00adtions plus inser\u00adtions plus sub\u00adsti\u00adtu\u00adtions, divid\u00aded by the num\u00adber of words in the human ref\u00ader\u00adence tran\u00adscript. Low\u00ader is bet\u00adter. But WER only means some\u00adthing when the ref\u00ader\u00adence tran\u00adscripts fol\u00adlow con\u00adsis\u00adtent rules.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">That is why tran\u00adscrip\u00adtion con\u00adven\u00adtions mat\u00adter. NIST\u2019s OpenASR21 plan tags hes\u00adi\u00adta\u00adtions, noise, over\u00adlap and for\u00adeign words, and treats some tokens as option\u00adal when scor\u00ading. Your style guide should set\u00adtle the same ques\u00adtions:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>How to write fillers, false starts and cut-off words.<\/li>\n\n\n<li>How to tag noise, laugh\u00adter, music, silence and over\u00adlap.<\/li>\n\n\n<li>Whether num\u00adbers, dates and acronyms are writ\u00adten as words or dig\u00adits.<\/li>\n\n\n<li>How to mark unclear speech and for\u00adeign words.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Beyond WER, track audio checks such as clip\u00adping, silence and chan\u00adnel errors. Add tran\u00adscript accu\u00adra\u00adcy on a dou\u00adble-reviewed sam\u00adple, label agree\u00adment between anno\u00adta\u00adtors, and meta\u00adda\u00adta com\u00adplete\u00adness. For more on tran\u00adscript work\u00adflows, see our guide to <a href=\"https:\/\/www.graveiensai.com\/blog\/ai-transcription\/\">AI tran\u00adscrip\u00adtion<\/a>.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Consent and compliance mistakes to avoid<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Voice is per\u00adson\u00adal data, and it can be bio\u00admet\u00adric data. The Euro\u00adpean Data Pro\u00adtec\u00adtion Board\u00ad\u2019s guide\u00adlines on vir\u00adtu\u00adal voice assis\u00adtants treat voice data used to iden\u00adti\u00adfy a per\u00adson as bio\u00admet\u00adric per\u00adson\u00adal data. When voice is processed to unique\u00adly iden\u00adti\u00adfy a per\u00adson, GDPR requires an Arti\u00adcle 9 con\u00addi\u00adtion as well as an Arti\u00adcle 6 legal basis.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">US law adds its own rules. As of Novem\u00adber 2026, the Illi\u00adnois Bio\u00admet\u00adric Infor\u00adma\u00adtion Pri\u00adva\u00adcy Act (BIPA) lists \u201cvoice\u00adprint\u201d as a bio\u00admet\u00adric iden\u00adti\u00adfi\u00ader. Cal\u00adi\u00adfor\u00adni\u00ada\u2019s CCPA defines bio\u00admet\u00adric infor\u00adma\u00adtion to include voice record\u00adings from which a voice\u00adprint can be extract\u00aded. This is not legal advice, so check your plans with coun\u00adsel.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Com\u00admon mis\u00adtakes we see:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Vague con\u00adsent.<\/strong> \u201cRecord\u00adings may be used to improve ser\u00advices\u201d does not cov\u00ader AI train\u00ading or syn\u00adthet\u00adic voic\u00ades.<\/li>\n\n\n<li><strong>No with\u00addraw\u00adal path.<\/strong> Speak\u00aders need a way to with\u00addraw, and you need a way to find and remove their files.<\/li>\n\n\n<li><strong>Reusing call-cen\u00adter logs blind\u00adly.<\/strong> Cus\u00adtomer calls were record\u00aded for anoth\u00ader pur\u00adpose. Check notice, con\u00adsent and local law first.<\/li>\n\n\n<li><strong>Cloning voic\u00ades with\u00adout a tal\u00adent agree\u00adment.<\/strong> TTS voice tal\u00adent should give explic\u00adit, spe\u00adcif\u00adic con\u00adsent for syn\u00adthet\u00adic use.<\/li>\n\n\n<li><strong>Weak meta\u00adda\u00adta.<\/strong> With\u00adout speak\u00ader IDs, demo\u00adgraph\u00adics and con\u00adsent links, you can\u00adnot prove cov\u00ader\u00adage or hon\u00ador dele\u00adtion requests.<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">How Graveiens AI approaches voice agent training data<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Graveiens AI col\u00adlects, tran\u00adscribes and checks speech data for con\u00adver\u00adsa\u00adtion\u00adal AI teams in the US and EU. Our <a href=\"https:\/\/www.graveiensai.com\/voice-speech\">voice and speech data<\/a> work fol\u00adlows an explic\u00adit-con\u00adsent work\u00adflow: par\u00adtic\u00adi\u00adpant onboard\u00ading, record\u00ading man\u00adage\u00adment, meta\u00adda\u00adta tag\u00adging and QA.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Lan\u00adguages.<\/strong> We cov\u00ader 25+ lan\u00adguages, includ\u00ading Indi\u00adan and Euro\u00adpean lan\u00adguages such as Ital\u00adian. See our notes on <a href=\"https:\/\/www.graveiensai.com\/blog\/indic-speech-data-collection\/\">Indic speech data col\u00adlec\u00adtion<\/a>.<\/li>\n\n\n<li><strong>Tran\u00adscrip\u00adtion.<\/strong> Our <a href=\"https:\/\/www.graveiensai.com\/transcription\">tran\u00adscrip\u00adtion<\/a> teams work to your style guide, with speak\u00ader turns, time\u00adstamps and event tags.<\/li>\n\n\n<li><strong>Qual\u00adi\u00adty.<\/strong> Graveiens AI is ISO 9001 cer\u00adti\u00adfied, and pilots are billed only on approved deliv\u00ader\u00adables.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">We do not promise mod\u00adel accu\u00adra\u00adcy gains. We give you con\u00adsent\u00aded, well-labeled data and the QA evi\u00addence to judge it.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Frequently asked questions<\/h2>\n\n\n<div id=\"rank-math-faq\" class=\"rank-math-block\">\n<div class=\"rank-math-list \">\n<div id=\"faq-question-1762000000000\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">What data do you need to train a voice agent?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>You need tran\u00adscribed con\u00adver\u00adsa\u00adtion\u00adal audio for speech recog\u00adni\u00adtion, labeled mul\u00adti-turn dialogs for the lan\u00adguage or dia\u00adlog lay\u00ader, and clean, con\u00adsent\u00aded record\u00adings if you build a cus\u00adtom TTS voice. The audio should match your live chan\u00adnel, accents and noise. You also need a held-out test set of real or real\u00adis\u00adtic calls from speak\u00aders not used in train\u00ading.<\/p>\n\n<\/div>\n<\/div>\n<div id=\"faq-question-1762000000001\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">How much speech data does a voice agent need?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>There is no uni\u00adver\u00adsal num\u00adber. It depends on your base mod\u00adel, lan\u00adguages, chan\u00adnel and how far your callers dif\u00adfer from its train\u00ading data. Most teams start with a focused pilot, mea\u00adsure results on a held-out test set from real callers, then expand cov\u00ader\u00adage where errors clus\u00adter. Speak\u00ader vari\u00adety often mat\u00adters as much as total hours.<\/p>\n\n<\/div>\n<\/div>\n<div id=\"faq-question-1762000000002\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">Is 8 kHz audio good enough for voice agent training?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>For phone agents, yes. 8 kHz nar\u00adrow\u00adband audio should be part of the mix because that is what callers send over clas\u00adsic phone lines. If your agent also runs in apps or on the web, add wide\u00adband audio too. Record at the native rate, keep each speak\u00ader on a sep\u00ada\u00adrate chan\u00adnel, and do not upsam\u00adple nar\u00adrow\u00adband audio.<\/p>\n\n<\/div>\n<\/div>\n<div id=\"faq-question-1762000000003\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">What is word error rate (WER)?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>Word error rate is the stan\u00addard accu\u00adra\u00adcy met\u00adric for speech recog\u00adni\u00adtion. Per NIST, it is the num\u00adber of dele\u00adtions, inser\u00adtions and sub\u00adsti\u00adtu\u00adtions in the sys\u00adtem out\u00adput, divid\u00aded by the num\u00adber of words in the human ref\u00ader\u00adence tran\u00adscript. Low\u00ader is bet\u00adter. WER is only com\u00adpa\u00adra\u00adble when both tests use the same tran\u00adscrip\u00adtion and nor\u00admal\u00adiza\u00adtion rules.<\/p>\n\n<\/div>\n<\/div>\n<div id=\"faq-question-1762000000004\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">Is voice data considered biometric data?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>It can be. The EDPB treats voice used to iden\u00adti\u00adfy a per\u00adson as bio\u00admet\u00adric per\u00adson\u00adal data, and GDPR adds stricter rules when voice is used to iden\u00adti\u00adfy some\u00adone. Illi\u00adnois BIPA lists voice\u00adprints as bio\u00admet\u00adric iden\u00adti\u00adfiers, and Cal\u00adi\u00adfor\u00adni\u00ada\u2019s CCPA cov\u00aders voice record\u00adings from which a voice\u00adprint can be extract\u00aded. Use explic\u00adit, doc\u00adu\u00adment\u00aded con\u00adsent and get legal advice for your case.<\/p>\n\n<\/div>\n<\/div>\n<div id=\"faq-question-1762000000005\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">Should voice agent data be scripted or spontaneous?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>Most projects use a blend. Role-play gives nat\u00adur\u00adal turn-tak\u00ading around real tasks. Script\u00aded prompts cov\u00ader hard enti\u00adties such as names, num\u00adbers and address\u00ades, and sup\u00adport TTS. Spon\u00adta\u00adneous con\u00adver\u00adsa\u00adtion adds real\u00adism and makes a strong test set. Set the mix based on where your agent fails today and what your live callers sound like.<\/p>\n\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n\n\n<h2 class=\"wp-block-heading\">Hear real conversational data before you commit<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Get a free 30-minute speech sam\u00adple set in your lan\u00adguage, accent and chan\u00adnel. Then scope a pilot billed only on approved files.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/www.graveiensai.com\/contact-us\">Request your free speech sam\u00adple set<\/a> or ask us for the Six-Axis Voice Cov\u00ader\u00adage Matrix check\u00adlist.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Sources<\/h2>\n\n\n\n<ol class=\"wp-block-list\">\n<li><a href=\"https:\/\/www.nist.gov\/document\/openasr21-challenge-evaluation-plan\" rel=\"noopener\" target=\"_blank\">NIST: OpenASR21 Chal\u00adlenge Eval\u00adu\u00ada\u00adtion Plan (WER def\u00adi\u00adn\u00adi\u00adtion, tele\u00adphone speech specs, nor\u00admal\u00adiza\u00adtion)<\/a><\/li>\n\n\n<li><a href=\"https:\/\/www.edpb.europa.eu\/system\/files\/2021-07\/edpb_guidelines_202102_on_vva_v2.0_adopted_en.pdf\" rel=\"noopener\" target=\"_blank\">EDPB: Guide\u00adlines 02\/2021 on vir\u00adtu\u00adal voice assis\u00adtants, ver\u00adsion 2.0<\/a><\/li>\n\n\n<li><a href=\"https:\/\/lig.ilga.gov\/legislation\/ilcs\/documents\/074000140K10.htm\" rel=\"noopener\" target=\"_blank\">Illi\u00adnois Gen\u00ader\u00adal Assem\u00adbly: 740 ILCS 14\/10, BIPA def\u00adi\u00adn\u00adi\u00adtions<\/a><\/li>\n\n\n<li><a href=\"https:\/\/leginfo.legislature.ca.gov\/faces\/codes_displaySection.xhtml?lawCode=CIV&amp;sectionNum=1798.140.\" rel=\"noopener\" target=\"_blank\">Cal\u00adi\u00adfor\u00adnia Leg\u00adisla\u00adtive Infor\u00adma\u00adtion: Civ\u00adil Code 1798.140, CCPA def\u00adi\u00adn\u00adi\u00adtions<\/a><\/li>\n<\/ol>\n\n\n\n<h3 class=\"wp-block-heading\">About the author<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Graveiens AI Edi\u00adto\u00adr\u00adi\u00adal Team.<\/strong> Our 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 run.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Voice agent train\u00ading data guide: ASR, dia\u00adlog and TTS needs, 8 kHz vs wide\u00adband, accents, turn-tak\u00ading, con\u00adsent and QA. Get a free speech sam\u00adple set.<\/p>\n","protected":false},"author":1,"featured_media":324,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"wp_typography_post_enhancements_disabled":false,"footnotes":""},"categories":[10],"tags":[45,44,46,43,42],"class_list":["post-325","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-voice-speech-data","tag-asr","tag-conversational-ai","tag-data-collection","tag-speech-data","tag-voice-agents"],"_links":{"self":[{"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/posts\/325","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=325"}],"version-history":[{"count":1,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/posts\/325\/revisions"}],"predecessor-version":[{"id":327,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/posts\/325\/revisions\/327"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/media\/324"}],"wp:attachment":[{"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/media?parent=325"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/categories?post=325"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/tags?post=325"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}