We collect and provide video training data for AI models to learn from.
Perit is a training-data company. Everything is recorded by the people who actually do the job, labelled by our own team, and licensed to you — including embodied data for robotics.
What we've delivered so far.
- 4,000+
- Hours of audio delivered
- 650
- Transcribers and aligners on the bench
- 50 hrs
- Delivered every working day
- 95%+
- QA-sample accuracy, checked against WER
Speech is our biggest line of work — 9 locales, audio delivered every working day, and a quality report with every batch. The same team, the same checks and the same reporting run our video, image and text work.
50 on fixed-term contracts, the rest freelance — and hiring against every open sector.
Freelance bench, timestamped to the word so a model can be scored on when it heard something, not just what.
Collect it, annotate it, or ask for the set that already exists.
Data collection
Speech delivered at scale. Video, image and text collected to spec.
Speech, video, image and text · nine locales
Data annotation
The Foundry bench — transcription, alignment and labels, any data type.
Audio, image, video and text · one calibration gate
Request data
Name the input your model fails on. We collect it, transcribe or label it, and license the set with the spec and terms published.
Priced by the unit it is delivered inFour steps, and the first one is a phone call.
No dataset catalogue to guess from. You describe the failure; we scope, pilot and deliver against it, with the contributor rate on the quote.
- 01day 0Tell us the failure
Not the dataset — the input your model gets wrong. A clip, a photo, a paragraph is enough to start.
- 0248 hoursWritten spec and quote
Quantity, locales, devices, conditions, annotation depth — and the contributor rate you will be paying, stated as its own line.
- 031–2 weeksPilot batch
Five percent of the volume, or ten to thirty episodes for embodied data, delivered first with a quality report. If the spec was wrong, the re-run is on us.
- 04ongoingWeekly delivery
Drops into your bucket in the region you chose, with the manifest, consent references and reviewer chain on every item.
Embodied data, collected by people who do the task.
Egocentric video, depth, teleoperation and handheld-gripper demonstrations, recorded in real homes, shops and workplaces across nine locales on the same recruiting, consent and QA machinery as our speech work.
Programs running in nine locales · quote in 48 hoursThis is a delivery, not a description of one.
Lowercase, fillers kept, one written convention for every tag, and a timestamp on every word — so accuracy and latency can be scored apart.
Read the methodhaan so i i filed the claim on monday [pause] but the portal is still showing pending only
let me pull that up for you ma'am [keyboard] can you confirm the claim reference
it's uh [pause] c l four two nine [overlap] eight—
[overlap] four two nine eight, yes i have it here
- Everything lowercase, punctuation only where it changes meaning.
- Fillers, repeats and false starts stay — they are what the model has to survive.
- [pause], [overlap], [keyboard] and other tags follow one written convention.
- Every word carries a start and end timestamp, so latency can be scored separately from accuracy.
Operators, not a crowd.
The bench is built from people who worked the floor, the ward and the claims desk — ex-contact-centre QA leads, claims handlers, ward clerks, shopkeepers. They are recruited for the sector or the locale a program needs, not pulled from a general pool.
Two crafts, one bench. Transcribers write what was actually said, fillers and all. Aligners put a start and end time on every word. A rubric author is someone who signs off on these calls at work, and every rubric has to be applicable by a stranger.
- Recruited for the sector
- The credential is checked at intake and attached to everything that person produces.
- Tested before they are paid
- A calibration set on every project, however long someone has been on the bench.
- Paid per approved unit
- The rate is on the listing before they start, and it is a line on your quote.
- Re-checked every week
- Reviewers are scored against the same gold set on a schedule, because graders drift too.
Crews on every continent, recruited to the locale you need.
Contributors record on their own phones, in their own rooms, in the accent and the light they actually live in. Blue is where the bench runs deepest today — 9 locales across Europe, Asia and the Americas — and anything beyond it is recruited to order against a locale quota.
Scored against the person who does the job.
Two speech benchmarks on held-out audio with a working operator as the control on every board, and live environments where the caller talks over your model.
Benchmarks
Two tracks — transcription and word alignment — with every model run on the same audio, under the same scoring, and ranked on the numbers.
Transcription · word alignmentRL Environments
Live speech-to-speech calls with trained speakers who interrupt, change their mind and hand over the phone mid-call.
Live in ten sectorsWhatever we collect, we pay the person who made it.
Pick a task on the workspace, do it on your own device in your own space, and get paid once it passes review. The rate is on the task before you start, and it is the same rate whoever the buyer is.
Twenty-minute sessions on your phone: read prompts, conversations, noisy rooms. Locales we are short on pay above base.
Five to fifteen minutes of a real task on a phone tripod, or on a chest camera we ship you. Rig programs pay per finished hour.
Shelves, products, documents and scenes to a brief, on the phone you already have. Matched-condition sets pay per pair.
Multi-turn dialogues, prompts and instructions in your own language, in the register of the sector you work in.
Independent, and staying that way.
No frontier lab owns a piece of Perit, and none of them sits on our cap table. Your held-out audio does not end up next to a competitor's. Backed by Y Combinator, F26.