What we can take on
Speech collection
Scripted, spontaneous, conversational, wake-word and in-car audio.
Image collection
Faces, documents, retail shelves, storefronts and field photography.
Video collection
Gestures, actions, in-cabin footage and 360 walkarounds.
Text and prompts
Written prompts, responses and dialogue produced to a taxonomy.
Sensor and device
IoT, accelerometer and LiDAR capture for on-device training.
Demographic quotas
Recruitment held to age, gender, region and accent targets you set.
The types of ai data collection we run
The full breakdown, not the headline categories. Anything not on this list is a scoping call away.
Speech. Scripted monologue reads, Spontaneous conversation, Wake-word and voice-command utterances, In-car and in-cabin speech, Call-centre speech by industry, Doctor dictation and consultation audio.
Image. Facial imagery by region and expression, Selfie and ID document pairs, Printed and handwritten OCR, Receipts, invoices and price tags, Retail shelf and storefront photography, Field and site photography.
Video. Gesture and action clips, Driver and in-cabin monitoring footage, 360-degree vehicle walkarounds, Screen recordings with intent labels, Public-scene and crowd footage, Aerial and drone footage.
Text. Prompts and instructions, Multi-turn dialogue scripts, Domain-specific writing samples, Parallel translation pairs, Structured form and survey responses, Classification and intent examples.
The forms the data takes
What you actually receive: file types, structure and the metadata that travels with it.
- Audio
- Mono or stereo WAV and FLAC at the sample rate you specify, with per-speaker channel separation where the recording allows it.
- Images
- JPG, PNG or HEIC at the resolution you set, with EXIF stripped or kept as you require, plus device and lighting metadata.
- Video
- MP4 clips at a set frame rate and resolution, single or multi-camera, with scene, lighting and subject metadata.
- Text
- JSONL or CSV with prompts, replies and dialogue turns, tagged by taxonomy, language and writer profile.
- Sensor streams
- Accelerometer, gyroscope, GPS and LiDAR logs captured alongside audio or video for on-device and mobility models.
- Metadata and consent
- A manifest per batch listing every file, the contributor profile fields you approved, device, environment and the consent reference.
How the data is collected
From the first brief to the batch you receive, step by step.
- 01
Recruited against your quota
Contributors are sourced from our community and vendor partners against the language, dialect, age, gender, region and device targets in your spec, not from whoever is online.
- 02
Captured on real devices
People record on their own phones, laptops or cameras, in the rooms, streets and vehicles your product will meet. Device model, OS and environment are tagged on every file.
- 03
Checked at the moment of capture
The recorder rejects clipped, silent, blurred or off-script takes before they upload, so a contributor retries in seconds instead of being rejected days later.
- 04
Consent recorded per file
Each contributor accepts a project-specific consent form. The record travels with the item, and a deletion request removes the file from every later delivery.
- 05
Reviewed by checks, then by people
Automated QC runs first, then peer review and expert review on a sample set by the risk of the task. Rejected items go back to the contributor with the reason.
How a project runs
- 01
Scoping call
We agree volume, languages, demographic quotas, formats and what counts as a rejected item.
- 02
Guidelines and gold set
Written guidelines with worked examples, plus a gold set every contributor is tested against.
- 03
Pilot batch
A small paid batch you accept before anything scales. Most guideline problems surface here.
- 04
Production
Work flows through capture checks, automated QC, peer review and expert review.
- 05
Delivery
Incremental batches in your format with a data card, QA report and agreement scores.
- 06
Iteration
Rejections feed back into the guidelines so the same mistake is not made twice.
How quality is checked
Checks that run during the work, not a sample at the end.
- Capture-time checks for noise, clipping, blur and length
- Native-speaker or domain review in every language
- Duplicate and near-duplicate detection across contributors
- Gold items hidden in the queue to measure contributors continuously
- A QA report with rejection reasons delivered with each batch
Industries that use it
Where teams bring us ai data collection, and what they use it for.
- AutomotiveIn-car speech, driver-monitoring video and road-scene imagery.
- HealthcareConsented clinical dictation and patient-conversation audio.
- Banking and financeCall-centre speech and financial document images.
- Retail and e-commerceShelf, product, receipt and storefront photography.
- TelecomSupport-call audio and network-field imagery.
- Real estateProperty photography and floor-plan capture.
- Safety and surveillanceScene and activity footage under stated conditions.
- Consumer devicesWake-word and voice-command audio from real homes and offices.
Why teams pick us for ai data collection
Quota recruitment
An open crowd call returns urban young men. Quotas cost more and are the only thing that works.
Consent per file
Every item carries a consent record and a deletion job that honours it.
Environment control
Room, device and noise floor specified and verified, not hoped for.
Capture-time checks
Bad audio is rejected in the browser before it ever uploads.
