# Modalities

## What we collect.

Four modalities, one machinery: recruit to a quota, consent on the file, a written protocol, a pilot batch, then weekly delivery with a quality report. What changes per modality is the kit and the protocol, not the people or the gate.

### Speech & audio

Two-party calls, spontaneous speech, read prompts, accents to quota, and the same script in a car, a kitchen and a warehouse.

- Two-party conversations, each side on its own channel
- Spontaneous and prompted speech with fillers kept
- Accent and locale coverage recruited to quota
- Matched acoustic conditions for degradation studies

4,000+ hours delivered · 9 locales live · 50 hours a day [Request speech data](/content/request-data/index.html)

### Video

Task video from a phone or an action camera — third-person and first-person — plus the embodied capture on the Physical Intelligence pages.

- Everyday tasks filmed in real homes and workplaces
- Egocentric capture from chest and head mounts
- Scripted scenarios with the same person across conditions
- Depth and gripper demonstrations, to a written protocol

Real homes and workplaces · 9 locales · phone, action-cam and rig capture [See Physical Intelligence](/content/physical-intelligence/index.html)

### Image

Photographs to a spec — products, scenes, documents, faces with consent — across the devices and lighting your users actually have.

- Product and shelf photography in real shops
- Scenes and interiors across locales
- Documents, receipts and forms, redacted to spec
- Same subject across devices and lighting

Products, scenes and documents · consent and redaction on every file

### Text & conversation

Written dialogues, prompts and chat transcripts authored by people who work in the sector, in the register the sector actually uses.

- Multi-turn dialogues to a scenario brief
- Prompts and instructions in nine locales
- Sector-specific written material by practitioners
- Transcripts paired with the audio they came from

Practitioner-authored · 9 locales · sector register, not crowd English

## How a collection runs

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

1. **Tell us the failure**  
Not the dataset — the input your model gets wrong. A clip, a photo, a paragraph is enough to start.  
2. **Spec and quote**  
Quantity, locales, devices, conditions, annotation depth — and the contributor rate you will be paying, stated as its own line.  
3. **Pilot batch**  
Five percent of the volume, delivered first, so the spec breaks early instead of at the end.  
4. **Weekly delivery**  
Drops into your bucket with a quality report and the reviewer chain attached to every batch.

### What arrives
What lands in your bucket.

- Original device capture, plus any agreed derivative
- Manifest per item: locale, device, condition, consent reference
- Transcript or labels to the agreed convention, where ordered
- Quality report per batch — accepted, rejected, and why
- Delivery into your region, on your bucket or ours

## Coverage

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, with the first crew on file inside two weeks.

- German (de_DE)
- French (fr_FR)
- Italian (it_IT)
- Spanish (es_ES)
- Portuguese (pt_BR)
- Hindi (hi_IN)
- Chinese (zh_CN)
- Korean (ko_KR)
- Japanese (ja_JP)

## Record, film, photograph, write

Whatever 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.

- Speech: $20 per approved audio hour
- Video: $6 per approved episode
- Image: $0.40 per approved photo
- Text: $8 per approved dialogue

Payout seven days after approval, $10 minimum balance, by bank transfer or UPI. Consent is on every file and can be withdrawn before payout.

## Questions

### Before you send a brief.

- Does the process change between speech, video, image and text?  
The machinery does not. Recruiting, consent, calibration and QA are the same for a photo as for a call; what changes is the protocol and the kit, which is what the pilot batch is for. The spec names both before anything is recorded.
- Who are the contributors?  
People recruited for the locale, the setting or the sector the spec calls for, paid per approved unit at a rate shown before they start. Anyone can join the speech pool; sector-specific work is recruited against the credential.
- How is consent handled?  
Captured on the file, before the contributor speaks or films, naming what the data trains. It can be withdrawn before payout, and the consent reference ships in the manifest of every item.
- Where does the data live?  
A collection is pinned to one region at kickoff — EU, US or India — and stays there. Graders stream, they do not copy.

## Not sure which modality you need?

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