AI training data collection
New text, speech, photos and video made to order by contributors you specify, in the languages and markets you need. Every item is checked, and every contributor signs a consent form that covers your use.
Voice prompt read aloud, 6.2 sC-0412, Monterrey, phone
AcceptedPhoto of a receipt, natural lightC-0388, Guadalajara
AcceptedWritten question about a phone billC-0451, Puebla
In reviewVoice prompt read aloud, 5.8 sC-0217, Mexico City, laptop
Redo: traffic noiseVideo of hand gestures, 12 sC-0398, Mérida, phone
AcceptedProduct review, 74 wordsC-0430, Tijuana
Accepted
- What we collect
- Text, speech, photos, video and documents
- Who contributes
- People matching the quotas you set by language, region, age and device
- Consent
- A signed form from every contributor, delivered with the data
- Ownership
- You own the data. We don't resell it as a dataset.
Data made for your model
Training data collection means creating new data your model needs and can't get elsewhere: questions asked the way your customers ask them, photos taken on the phones they use, speech in the accents they have. We recruit contributors to your spec, brief them, check every item and record their consent.
We collect to order only. We don't sell off-the-shelf datasets, so nobody else gets the data you paid for.
What we collect
Most projects pair collection with labeling, so the data arrives ready to train on.
Text and prompts
Questions, requests, reviews and chat messages written by native speakers in their own words, to topics and intents you define.
Speech
Read prompts, free speech and conversations to your speaker mix. See speech and audio data.
Photos
Objects, scenes, receipts and products photographed to a shot list, on the devices and in the lighting you specify.
Video
Short clips of actions, gestures or tasks, filmed by contributors who match your demographic quotas.
Documents
Filled-in forms, letters and handwriting samples in many scripts, created with made-up details so no real personal data is involved.
Smaller languages
Data in languages that are hard to find online, such as Tigrinya, Somali, Pashto or Karen, from contributors we vet for each language.
How a collection project runs
You see real items from the pilot before we recruit at scale.
Spec
Data type, languages, quotas, devices, file specs and the use the consent form must cover.
Recruit and brief
We find contributors who match, explain the task with examples and collect signed consent.
Pilot review
You review a pilot set. We adjust the brief and the checks before full collection.
Your sign-off QACollect and check
Every item is checked against the brief. Items that fail go back for a redo.
Delivery
Data, metadata and consent records in batches, with quota progress in each report.
What we check on every item
Items that fail any check go back to the contributor or are replaced.
- Brief matchThe item does what the task asked, in the right language and variety.
- Technical specsResolution, length, sample rate, file type and naming as specified.
- DuplicatesNo repeated or near-identical items from the same contributor.
- Personal dataNo real names, numbers or faces of bystanders unless your consent setup covers them.
- Quota fitContributor attributes recorded, so each item counts toward the right quota.
Consent and ownership
Set out in the contract and in each contributor's consent form. See how we handle data.
- Consent for your use. Every contributor signs a form naming how the data will be used. You get the signed records.
- Contributor IDs, not names. Deliverables carry an anonymous ID and the attributes you need.
- Yours alone. You own the collected data, and we don't sell it or reuse it for anyone else.
- Paid contributors. Every contributor is paid for accepted work.
Formats
- Text
- JSONL, CSV or TXT, with one item per record and its metadata
- Audio
- WAV or FLAC at the sample rate and bit depth you set
- Images and video
- JPG, PNG, HEIC, MP4 or MOV at your minimum resolution
- Metadata and consent
- Contributor ID, attributes, device and setting in CSV or JSON, plus signed consent forms
Recruiting setup and contributor payment terms to confirm
Related services
Questions about data collection
Do you collect data to order or sell existing datasets?
We collect to order only. Each project is built to your spec, and the data isn't sold to anyone else.
Who owns the data?
You do. The contract assigns ownership of the collected data to you, and each contributor's consent form covers the use you describe.
How do you record consent?
Every contributor signs a consent form before taking part. It names the purpose and how the data will be used, and you receive the signed records with the data.
Can you collect data in smaller languages?
Yes, including languages with little data online. For each language, we confirm we can recruit enough contributors before the pilot starts, so there are no surprises halfway through.
How do you check what contributors send?
Every item is checked against the brief, the technical specs and our personal-data rules. Items that fail go back for a redo or are replaced, and the batch report shows acceptance rates.
Scope your collection
Send your data type, languages and quotas, or book a call. We propose a pilot collection and a price.
- Phone
- +44 20 3290 0116
- Phone hours
- Weekdays 9:30 to 16:00, weekends 9:30 to 12:00, UK time