Your model learns from whoever is in its training data. So our data collection starts with your users: the languages and accents they speak, the phones in their hands, the streets and shops their cameras see. Then people from our 30,000+ network collect the speech, text, images and video to match, and we check it before you get it.
"I'd like to block my debit card."
Data collection is the work of gathering the raw speech, text, images and video a model learns from, matched to the languages, accents and settings the model will actually meet once it ships.
In practice that means recording someone reading a prompt or taking a call, writing everyday messages in a banking app, photographing a shelf in real store lighting, or filming a hand gesture from the angle a camera will use. The data is only as useful as the fit between who supplied it and who the model will serve.
Collecting from a general pool is faster, but it rarely matches a specific product's users. For what to check before you commit to a dataset, see speech data and image data. For cases where collection isn't the right fit, our guide to synthetic data vs real data covers the alternative.
Tell us what your model has to handle and we'll collect for it. Speech, images and video run through collection tools we built ourselves, so the data can go straight on to labeling or transcription without another vendor. Each one has its own page with a sample: custom speech data collection, image data, video data and text data.
First-person footage from head mounts, action cameras and smart glasses is on the egocentric video data page.Read prompts, free conversation and call audio, recorded by speakers chosen to match your users in language, accent, age and region.
For speech recognition, voice assistants, voice bots and IVR
Our tools: Prompt Audio and Conversation Audio Free sample of speech and audio data
Prompts, questions, answers and everyday messages, written by people in the languages and the domain you need.
For chat assistants, search and your language models
Photos of products, shelves, streets and documents, taken on the kinds of phones and in the kind of light your model will meet. Our latest project was 16,000 images from 2,400 participants.
For object detection, recognition and shelf checks
Our tool: Image Collection Free sample of image data
Clips of actions, gestures, driving and everyday scenes, filmed to your brief for angle, setting and length.
For action recognition, video understanding and driver assistance
Our tool: Video Collection Free sample of video data
Paired data, like a photo with a spoken description and a written caption, collected together so every part matches.
For models that see, hear and read at once
Our tool: Media Collection Free sample of multimodal data
Material from inside your company, like support tickets, forms and call recordings, gathered for an internal AI project under an NDA.
For internal assistants and tools built on your own material
The samples on these cards are illustrations, not client data.
Common jobs for freshly collected data, and the kind of collection behind each one. They're examples of the work, not a client list.
Recorded prompts and calls teach a voice assistant to recognize how your users actually talk, not just how a script reads.
Everyday messages written in your domain give a chat assistant real phrasing to learn from, instead of textbook sentences.
Shelf and product photos, taken on the kinds of phones your team uses, teach a model to recognize what's actually on the shelf.
Clips filmed to your brief teach a model to recognize actions, gestures or road scenes the way your own cameras will see them.
A photo, a spoken description and a caption collected together teach a model to work across all three at the same time.
Support tickets, forms and call recordings gathered under an NDA train a tool on your company's own language, not a generic one.
A few choices at the start save a lot of rework later. These are the ones we'd settle with you first.
Age, gender, region and accent all shape what the data sounds like. A dataset that only captures one kind of speaker only works for that kind of speaker.
Scripts give you control over coverage. Spontaneous speech brings the hesitations and interruptions real users produce, and many projects need both.
A quiet room, a busy street, a phone line, the microphone your users hold. Match the recordings to that, or your model meets sounds in the real world it never heard in training.
Decide the training and test split before collection starts, and keep the test part away from training. Settle what contributors agree to, and what the data may be used for, at the same time.
Existing datasets are a fine place to start, and for some problems they're all you need. They run out when your product needs a language, an accent, a domain or a recording condition that nobody has collected at the size you need.
Take a call center bot for a regional bank. Its callers use local words, switch languages mid-sentence and phone in from noisy places. General speech data doesn't sound like that. Data collected for the job can.
If generated data is also on the table, our guide to synthetic data vs real data covers when each one makes sense.
One named person owns the project from scoping to the last file. This is the order things happen in.
You tell us the domain, the languages and roughly how much you need.
We sign an NDA before any data moves, then agree the scope with you.
We collect a small sample free, so you can check it before you commit.
People from our network collect in our tools, in the languages and conditions you asked for.
A review pass runs before anything is delivered.
The data comes to you, or goes straight on to labeling, transcription or translation with us.
Give us the language and the accent or dialect, and we'll find contributors who speak it. In India we go well past Hindi and English. Our page on Indian language data explains how.
The words that come up when you plan a collection project, and what each one means.
We run labeling, transcription and translation ourselves, so collected data moves on without going to another vendor. You can also compare all eight services. Example category: automotive driving scene video.
Tags, boxes and segmentation added by people to the data we collected.
Collected speech and video turned into text, with a review pass built in.
Prompts and text carried into the other languages your model has to handle.
Scripted and natural dialogue for voice bots, IVR and support lines.
Generated data, reviewed by people, for cases where real data is scarce or sensitive.
Short answers about collecting data with us. The services FAQ covers the other services and how they fit together.
Speech and audio, text, images, video, mixed-media and multimodal data, and corporate data for internal AI projects.
People from our network of 30,000+ worldwide. We pick contributors who fit your brief: the language, the accent or dialect, and the kind of person and setting your model has to handle.
Yes. We work in languages from all over the world, with deep coverage of Indian ones well past Hindi and English. Give us the language and the accent or dialect and we'll find contributors who speak it.
Yes. Send a short brief with the data type, the languages and a rough volume, and we'll collect a small sample free before you commit.
Yes. We run labeling, transcription and translation ourselves, so collected data can move straight on without going to another vendor.
No. We use it only for the scope we agreed. It's never resold or reused on unrelated projects, and we don't train models of our own with it. Your team trains your model.
Yes. We return it or delete it when you ask, and until then only the people working on your project can get to it.
Pick the kinds of data, add the languages and a rough volume, and send it. We'll collect a small sample free before you commit.
This opens your email app with the brief filled in. Nothing is sent until you press send. You can also use the contact page.