Search for voice data and you will find broad data-service providers, speech-focused firms, dataset marketplaces and open repositories. Those labels do not prove source quality, language reach, rights or the ability to deliver. This guide turns names such as Appen, Defined.ai, Shaip, TELUS Digital, Sama, Sigma.ai, Spirelight and Mozilla Common Voice into a shortlist to check, not a ranking.
Owners and services change. Check current official evidence, and compare every candidate against one written brief. It should cover speakers, conditions, annotations, provenance, rights, QA, sample evidence, capacity, schedule and price.
Who provides voice training data
Buyers usually meet four kinds of supplier: broad managed providers, speech-focused providers, dataset marketplaces and open repositories. A marketplace may list data from many publishers. A managed supplier may collect the data itself or through partners. Some providers fit more than one category.
Names buyers may consider include Appen, TELUS Digital, Sama, Defined.ai, Shaip, Sigma.ai, Spirelight and Mozilla Common Voice. A name on this list makes no claim about its current capability, quality, ownership or fit. Check current official materials, the exact publisher or collection partner, and a proposal for your project. The Appen alternatives guide and AI training data company guide give related buying frameworks.
Seven fields to compare
| Field | Evidence to request |
|---|---|
| Language and dialect | Target variety, recruiting channel, screening method, speaker cells, subcontractors, and feasibility assumptions |
| Recording conditions | Device, microphone geometry, environment, channel layout, speech type, and deployment match |
| Volume and capacity | Usable-hour definition, pilot size, recruiting capacity, dependencies, schedule, and rejection assumptions |
| Collection status | Finished inventory, third-party listing, custom collection, or feasibility target; version and sample availability |
| Rights and privacy | Publisher and source chain, permissions, license, applicable lawful basis and notices, retention, transfers, and consent records when relied on |
| QA and annotations | Guidelines, reviewer qualifications, agreement method, sampling, thresholds, rework, and sample evidence |
| Engagement and commercials | Named roles, subcontractors, change control, security, minimums, included work, currency, taxes, schedule, and final price |
The supplier's category does not predict the answers. For the exact delivery, check the publisher, source chain, permissions, speaker coverage, subcontractors, recording method and QA. Do not infer them from the category.
Rights, privacy, and provenance
Start with the intended use, not a one-size consent checklist. Ask what rights the supplier or publisher actually holds, what the license permits and which privacy records apply. For personal data, the lawful basis, notices, roles, retention, transfers and data-subject handling depend on the project. If consent is relied on, ask for the evidence and the withdrawal process.
Exclusivity, redistribution, model-related rights, territory, term, retention and risk allocation depend on the contract. The speech data licensing guide explains the questions to put in the diligence record.
Low-resource and multilingual sourcing
A headline count of languages does not prove dialect coverage or the capacity to recruit. Ask for the named variety, where speakers are recruited, how they are screened, the target speaker cells and the reviewers' qualifications. Ask too for subcontractor disclosure, pilot evidence and schedule assumptions. See the low-resource language guide for a detailed brief.
Finished license or custom collection
A finished asset can shorten lead time when its source, speakers, conditions, annotations, rights and QA match the use. A custom collection may fit a narrower brief, but its feasibility, pilot design, capacity, schedule and price still need confirming.
Compare both routes against the same acceptance test. Neither one is faster or cheaper by default. The custom collection guide covers the workstreams to specify.
Where Spirelight may fit
Spirelight can assess a voice-data brief. The sourcing, protocols, reviewers, permissions, privacy and provenance records, license terms, annotations, QA, sample evidence, capacity, schedule and price it needs must then be confirmed in writing. Its dataset pages state each dataset's hours, speakers and price per hour; the documentation a specific project needs is still agreed in writing.
Send the languages, speaker cells, recording conditions, intended use, volume and acceptance criteria through the written project brief. The reply will say whether the scope can be assessed and what evidence is available.