Request data — Perit

13 things, priced by the unit they are delivered in.

Grouped by modality. Every line is priced by the unit it is delivered in, with its lead time on the card and the contributor rate on the quote.

Speech & audio

Two-party conversations 2–4 weeks
Recorded calls between a caller and an agent, both sides on separate channels.

You receive Stereo audio + verbatim transcript + turn boundaries Priced per hour of usable audio

Spontaneous speech 2–3 weeks
Unscripted speech from real speakers — the hesitations, restarts and overlaps scripts never produce.

You receive Mono audio + verbatim transcript with disfluency markup Priced per hour of usable audio

Accent & language coverage 3–5 weeks
Targeted read and prompted speech to fill the accents your model drops.

You receive Audio + transcript + speaker profile Priced per speaker (30–60 min)

Real acoustic conditions 2–4 weeks
The same script recorded in a car, a warehouse, a call centre and a kitchen.

You receive Parallel audio per condition + condition labels Priced per condition-hour

Transcription & diarization 3–10 days
Your audio, transcribed and speaker-attributed to a written standard.

You receive Transcript, speaker turns, timestamps, agreement report Priced per audio hour

Entity & intent labeling 1–2 weeks
The tokens that carry the transaction — account numbers, amounts, drug names, SKUs, intents.

You receive Span labels + intent tags + schema Priced per audio hour

Preference & rating sets 1–2 weeks
Side-by-side judgments on model replies: which one a customer would accept.

You receive Pairwise preferences, rubric scores, MOS/MUSHRA where relevant Priced per judgment

Adversarial & red-team audio 2–4 weeks
Callers who lie, push, spoof and hand the phone to someone else mid-call.

You receive Attack audio + severity labels + expected refusal Priced per engagement

Video

Egocentric task video pilot in 2 weeks
First-person video of people doing everyday tasks in real rooms, from a chest or head mount.

You receive Raw video + task and environment metadata + QC status Priced per finished hour

Image

Product & scene imagery 2–4 weeks
Photographs to a spec — shelves, interiors, objects — on the devices and in the light your users actually have.

You receive Images + capture metadata + labels where ordered Priced per image

Documents & forms 2–3 weeks
Receipts, forms and IDs photographed the way people actually photograph them, redacted to spec.

You receive Images + field extraction where ordered Priced per image

Text & conversation

Written dialogues & prompts 1–3 weeks
Multi-turn conversations and instructions authored by people who work in the sector, in its register.

You receive Dialogues in JSON + author metadata Priced per dialogue

Conversation transcripts 1–3 weeks
Transcripts paired with the audio they came from, or written to a scenario where no audio exists.

You receive Transcript + turn boundaries + optional audio pairing Priced per conversation

How a request runs

Four steps, and the first one is a phone call.

  1. Tell us the failure
    Not the dataset — the call your model gets wrong. A clip is enough to start.
  2. We scope and quote
    Quantity, locales, conditions, annotation schema, and the contributor rate you will be paying.
  3. Pilot batch
    5% of the volume, delivered first, so the schema breaks early instead of at the end.
  4. Delivery
    Weekly drops into your bucket, each with an agreement report and the reviewer chain attached.

What arrives

The delivery spec, published.

These are the defaults for speech, and the pattern for everything else. Every one of them is negotiable in the written spec we agree before the pilot — but nobody should have to ask what a delivery contains.

Audio 16-bit PCM WAV at 16 kHz by default, 48 kHz on request. Mono, or two-channel where both sides were recorded separately. Original device capture is kept alongside anything we resample.

Transcript UTF-8 JSON per clip, with a plain-text and CSV mirror. Verbatim by default: fillers, repeats and false starts stay, and [pause], [overlap] and noise tags follow one written convention agreed in the spec.

Alignment Start and end timestamps on every word, so accuracy and latency can be scored apart. Speaker turns and diarization labels ride in the same file.

Manifest One row per clip: duration, locale, speaker profile, device, acoustic condition, consent reference and the reviewer chain that signed it off.

Quality report Per batch: QA-sample accuracy, WER against the adjudicated reference, inter-annotator agreement on the double-passed portion, rework rate, and what was rejected and why.

Handover Weekly drops to your S3 or GCS bucket, or ours with time-boxed credentials. Data stays in the region agreed at kickoff.

Before you ask

Ownership, consent, residency.

Who owns what we deliver?
You get a perpetual, non-exclusive licence for internal model development and evaluation. Exclusive and time-limited-exclusive terms exist — ask, and they are priced on the order form.

How was it consented?

Every contributor consents on the recording, before they speak, to exactly this use. Consent is stored with the file and referenced in the manifest, and it can be withdrawn — which is why the consent is taken up front rather than after.

Where does it live?

A collection is pinned to one region at kickoff — EU, US or India — and the audio, transcripts and labels stay in it. Graders stream the clip rather than copy it.

What if the pilot is wrong?

The pilot batch is 5% of the volume and it exists to break the schema early. If it comes back wrong, we fix the spec and re-run it before the main volume starts — that is what the pilot is for.

Not sure which of these you need?

Send one input your model handles badly. We will tell you what would fix it, and what it would cost — before anyone signs anything.