If you need to buy AI training data, the short version is this: decide first whether you truly need to buy (versus building in-house or using open data), then match your need to the right provider type (marketplace, BPO, or specialist vendor), and evaluate every candidate on four things: consent chain, license scope, QA evidence, and a sample audit you run yourself before signing anything.
Most training-data purchases go wrong at one of two points. Either the buyer picks the wrong provider category (paying custom-collection prices for something that exists off the shelf, or expecting a marketplace to deliver studio-grade quality), or they skip legal diligence and discover post-training that the data was scraped, resold, or never consented for commercial ML use. This playbook walks through both problems, with the specific questions to ask and the red flags that should end a conversation.
We sell speech datasets ourselves, so we have a horse in this race. The framework below applies to any modality, though: text, image, video, or audio.
Buy, build, or scrape? Decide this first
Before comparing training data providers, decide whether purchasing is the right move at all.
Build (collect in-house) when the data is proprietary to your product: user interactions, domain documents, telemetry. Nobody can sell you your own users' behavior, and it's usually your strongest moat. The hidden cost is tooling, consent management, and annotation infrastructure, which routinely takes 3-6 months to stand up.
Use open data when it genuinely covers your case. Common Voice, LibriSpeech, LAION, and similar corpora are free, but check licenses carefully: many open datasets are research-only (CC BY-NC) or carry share-alike terms that are awkward for commercial models. Open speech data in particular skews toward read/prompted English.
Scrape only with a lawyer in the room. Post-2024, the litigation and regulatory environment around scraped training data (copyright suits, GDPR/BIPA claims for voice and biometric data, the EU AI Act's transparency requirements) has made "we found it on the internet" an expensive position. Scraping can be defensible for some public text; for voices, faces, and anything personal, it usually is not.
Buy when the data exists, is legally clean, and building it yourself would cost more than the license. For a 1,000-hour conversational speech corpus, in-house collection typically means recruiting 100+ speakers, building consent workflows, QA-ing audio, and transcribing. Buying the finished dataset is almost always cheaper and 6-12 months faster.
Types of training data providers
There are three broad categories, and mixing them up is the most common procurement mistake.
| Provider type | Examples of the model | Best for | Typical weakness |
|---|---|---|---|
| Crowdsourcing marketplaces | Task platforms with large distributed workforces | High-volume, low-complexity labeling; RLHF at scale | Variable quality; thin consent documentation; you manage the spec |
| BPO / labeling firms | Managed annotation teams, often offshore | Ongoing annotation pipelines, complex labeling with training | Slower ramp; per-seat pricing; data collection (vs. labeling) often outsourced again |
| Specialist dataset vendors | Domain-focused shops selling ready-made or semi-custom datasets | Speech, medical imaging, LiDAR, low-resource languages | Narrower catalogs; you buy what exists unless you fund custom collection |
A useful rule: marketplaces sell labor, BPOs sell managed labor, specialists sell finished data. If you need 800 hours of transcribed Egyptian Arabic conversation next month, a specialist with an existing corpus (for example, an off-the-shelf Egyptian Arabic dataset) will beat any labor-based model on both price and time, because collection already happened. If you need 2 million bounding boxes drawn to your ontology, the reverse is true.
For needs no catalog covers (a specific dialect, acoustic condition, or domain vocabulary), specialist vendors usually also offer custom speech data collection, which prices higher but delivers exactly your spec.
Evaluation criteria: the four checks that matter
1. Consent chain
Ask: "Show me what contributors signed." A legitimate vendor can produce the actual consent language granting rights for commercial AI training, and can map every file in the dataset to a consenting contributor. For speech and image data this is non-negotiable, since voices and faces are biometric identifiers under GDPR and several US state laws. If the vendor's answer is "the data is publicly available," that is not consent; that is a scraping disclosure.
2. License scope
Read the license for five things: (a) commercial model training explicitly permitted, (b) whether trained models can be deployed commercially (some licenses cover training but not deployment, yes, really), (c) sublicensing and cloud-provider terms if you train on rented infrastructure, (d) perpetuity vs. subscription terms, and (e) what happens to your model if the license terminates. Our speech data licensing guide covers the clauses that bite most often.
3. QA evidence
Don't accept "99% accuracy" as a bare claim. Ask how it was measured: inter-annotator agreement scores, transcription error rates on audited samples, the review workflow (single-pass vs. double-blind), and rejection rates during collection. Vendors who run real QA can answer in specifics within a day; vendors who don't will send you a marketing PDF.
4. Your own sample audit
Before any contract, get a free or cheap sample (1-5% of the dataset, selected by you, not curated by the vendor) and audit it against your acceptance criteria. For speech: listen to random files, check transcripts word-by-word on a subsample, verify metadata (speaker IDs, demographics, sample rates), and run it through your existing model to see if it behaves like the distribution you expect. An hour of engineer time here saves five-figure mistakes.
Pricing models you'll encounter
- Per unit (per audio hour, per image, per 1k tokens): standard for off-the-shelf data. Conversational transcribed speech generally lands at $60-150/hour off the shelf; our catalog runs $60-95/hour depending on language.
- Per labeler-hour or per seat: standard for BPOs. Watch for minimum commitments.
- Project-based: standard for custom collection. Expect 30-50% deposits and milestone payments.
- Subscription/API access: common for continuously updated data. Check what happens to models trained during the subscription after you cancel.
Negotiation levers that actually work: volume tiers across languages or datasets, exclusivity carve-outs (non-exclusive licenses should be meaningfully cheaper), and paying for a pilot tranche with pre-agreed pricing on the remainder.
Red flags that should end the conversation
- No consent documentation. The polite term for consented-sounding data with no paper trail is "data laundering": scraped audio or images passed through an intermediary who adds a license they had no right to grant. You inherit the liability, not the intermediary.
- Resold public datasets. Some resellers package LibriSpeech, Common Voice, or YouTube-derived audio and sell it as proprietary. Spot-check samples against known corpora; ask pointed questions about collection dates, locations, and recruitment method.
- Prices too good to be true. Genuine consented collection has a real cost floor; recruiting, paying, and QA-ing contributors isn't free. Transcribed speech at $10/hour was scraped, full stop.
- Refusal to allow a sample audit or insistence on curated "showcase" samples only.
- Vague provenance answers. "Our partners collect it" without naming the method, geography, or consent process.
- No warranties or indemnification in the contract. Vendors confident in their provenance will stand behind it contractually.
Procurement checklist
Copy this into your RFP:
- Written description of collection methodology (who, where, when, how recruited, how paid).
- Copy of contributor consent language; confirmation every file maps to a consent record.
- License draft reviewed for: commercial training, commercial deployment, sublicensing, term, termination effects on trained models.
- QA methodology with measurable metrics (IAA, audited error rates) and rejection statistics.
- Buyer-selected sample (≥1% of corpus) delivered before contract; audit passed against written acceptance criteria.
- Metadata completeness check: speaker/contributor IDs, demographics, device/channel info, sample rates or resolutions.
- Overlap check against major open datasets.
- Warranties on provenance and non-infringement; indemnification clause; liability cap negotiated.
- Delivery format, re-delivery terms, and support SLA in writing.
- Data protection terms (GDPR/CCPA processor language) if any personal data is included.
For a speech-specific version of this process, see our deeper guide on how to buy speech data.
Need training data?
SpeechData.ai offers off-the-shelf conversational speech datasets in 60 languages: 500-2,000 hours each, fully transcribed, consent-documented, and licensed for commercial ASR, TTS, and voice-AI training. Browse the dataset catalog or contact us for samples and licensing details.