AI voice training jobs and video data creation explained

What voice and video data creation involves

Data creation requires new material to be produced rather than found. Nothing is being gathered from the internet or from an archive. Contributors record something new, following strict specifications, because the specification describes conditions that no existing recording matches.

A project may need a particular accent, spoken at a particular pace, in a room with a particular level of background noise. Another may need the same household task filmed by fifty different people in fifty different kitchens. The variation is the product, which is why briefs read more like technical instructions than creative ones.

Polish is not the goal and can actually work against you. A professionally produced, studio-clean recording of a scripted sentence is less useful to a speech model than an ordinary phone recording made in a real kitchen because the second sounds more like real-life conditions AI models will actually encounter. Following the specification exactly matters far more than sounding or looking good.

Types of audio recording projects

Audio is the most established form of contributor data creation, and the task types are more varied than "read some sentences" suggests.

• Scripted reading: You read a fixed list of sentences, phrases, or numbers into your phone. The most common entry-level task type, and the easiest to check before submitting.

• Spontaneous speech: You speak naturally on a given topic rather than reading. Projects use this because scripted speech has different rhythm and pronunciation patterns from real conversation.

• Wake words and commands: Short phrases of the kind people use with voice assistants, often repeated at different distances, volumes, and speeds.

• Guided dialogue: Two or more speakers hold a conversation following a scenario. These usually pay more per session because they are harder to schedule and coordinate.

• Targeted accent and dialect recording: Projects recruit for a specific language, regional accent, dialect, age group, or speaker profile, because coverage gaps are the reason the dataset is being built.

The environment is often specified alongside the speech itself — maybe a quiet room, outdoors, with traffic noise, or inside a vehicle. Recording in the wrong environment leads to review failure even when the speech is perfect, so read that part of the brief before choosing where to sit.

Accent and language coverage is the reason so many of these projects exist. Open initiatives such as Mozilla Common Voice were founded on the same observation, that speech systems perform worse for speakers whose voices are underrepresented in training data. This is also the reason for a growing number of projects seeking bilingual or multilingual speakers, fluent in widespread languages like Spanish and Chinese. The goal is to record how speakers naturally sound, sometimes weaving two or more languages together.

Looking for voice recording projects? Explore our open opportunities here!

Types of video recording projects

Video projects ask for movement, sequence, and context, which makes their specifications tighter than audio.

• Everyday activities at home: Cooking, cleaning, laundry, small repairs, organizing, plant care, and pet care. The most widely available category because almost everyone has access to the setting.

• Hands-on work activities: Automotive work, car washing and detailing, construction, farming, plumbing, electrical work, woodworking, manufacturing, food service, hospitality, warehouse work, and recycling. These are harder to find and sometimes better paid.

• Gestures and demonstrations: Performing a defined action or sequence, sometimes repeated with different objects or from different angles to create usable variation.

• First-person footage: Recorded from your own point of view with the phone on a head mount, so your hands stay free and the camera sees what you see. Read our guide to video data collection projects for more details.

Repetition is often part of the assignment. A project asking for the same task filmed five times with different objects wants exactly that, and submitting one take with four near-identical copies is a common reason for rejection.

What makes a recording usable (and what gets it rejected)

Review decisions in this category are mechanical rather than subjective, which is good news – the criteria are very clear in advance.

For audio, submissions typically fail on: 

• Missing or misread phrases

• Background noise above the specified level

• Clipping from speaking too close to the microphone

• Volume too low to be usable

• The wrong file format

Most of these are visible in a quick playback before you upload, and checking one recording takes less time than redoing a batch.

For video, common causes for rejection are: 

• The required subject leaving the frame

• Poor lighting

• Unstable camera work

• An incomplete action

• Footage shorter than the stated minimum

• Near-duplicate clips

The single most common failure on first-person projects is hands drifting out of view during the part of the task that matters.

To avoid rejection, two good practices to incorporate into your session are:

• Record one short sample first: Submit it, wait for the verdict, then produce the rest. Contributors who record a full batch before their first review result are the ones who lose hours to a fixable misreading of the brief.

• Read the rejection criteria section specifically: When a brief states what counts as unacceptable, treat that section as the definition of the task. It is more useful than the description of what to do.

Payment applies to accepted submissions, so a high early rejection rate quietly cuts your effective hourly rate even when the published one looks fine.

Setting up so your first submission passes

Setup takes ten minutes the first time and two minutes afterward. It is the highest-value part of the task. Use this checklist to get started: 

▢ Check the device requirement before anything else: Projects specify phone models, operating systems, camera modes, or file formats, and an unsupported device fails automatically regardless of what you recorded.

▢ Find the quietest usable room for audio: Soft furnishings help, hard empty rooms echo. Fridges, fans, and air conditioning are audible on recordings even when you have stopped noticing them.

▢ Hold the phone a hand's width from your mouth: Closer causes clipping and breath noise, further drops the level below what the brief asks for.

▢ Use the widest camera setting for video: Most first-person projects need the ultrawide lens so hands and workspace fit in frame together, but check the guidelines to be sure.

▢ Record in daylight where possible: Kitchens and garages are the two settings where lighting most often fails review.

▢ Watch the sample back before submitting: Thirty seconds of checking prevents most rejections in this category.

Recording outside your home and the extra rules that apply

Footage recorded in public spaces is among the most valuable in this category, because everyday outdoor settings such as parks, streets, and shops are underrepresented in training data. It also comes with responsibilities that recording at home doesn't: the people around you haven't agreed to appear in your video. 

Before you start, choose a time and spot with little foot traffic, such as an empty corner of a park early in the morning rather than a busy path at midday. Frame your shot so it excludes other people, license plates, house numbers, and screens. If someone walks into view, pause the recording and wait until they've left, or delete the clip. Never record children, and follow any posted rules or local laws that restrict filming.

A public space that’s highly sought after in recording projects is the workplace, especially industrial and trade setting. Four rules typically apply before you film anywhere you work:

• It has to be a genuine workplace: A home space arranged to resemble one does not qualify, and staged setups are identifiable in review.

• Your manager should know: If you are not the owner of the business, tell whoever is responsible before recording anything.

• Safety equipment comes first: Wear your safety gear as usual and place the camera around it. Never move or remove protective equipment for the sake of footage.

• Keep other people out of frame: Coworkers, customers, faces, badges, screens, and documents should not appear. 

You generally can't record at client sites, customer premises, or anywhere with confidential materials. When a setting is ambiguous, the brief's guidance is the answer, but double-checking with the project manager is always a good idea. 

What you need to participate and how payment works

To participate in a recording project, you generally need: 

• An eligible location

• A supported device

• The requested speaker or participant profile (where one applies)

• A suitable recording environment

• Precision in following instructions

Eligibility is often set by location because datasets need contributors from specific populations and because consent rules differ by jurisdiction. We publish our current eligibility and geographic restrictions separately from individual projects. 

Payments vary by project and platform. You might be paid per:

• Accepted hour

• Recording

• Session

• Approved batch

• Completed task

Privacy when your voice, face, or workplace is in frame

Voice and video recordings carry more personal information than most task-based projects. Follow these basic safety steps to protect yourself before committing to a project:

• Read the consent and privacy terms before recording: Establish what the material may be used for, how long it is kept, who can access it, and whether deletion can be requested. In some countries, voice and face recordings count as biometric data, so extra consent steps are normal, not a red flag.

• Before you press record, clear the frame: Remove all documents, screens, mail, medication, keys, and anything showing an address or full name. 

• Keep other people out of shot entirely: Unless the brief explicitly permits it and consent has been obtained. Keep in mind, children cannot meaningfully consent to their image being used as training data. 

• Check that background conversation is not audible: Other people's voices need the same care as their faces.

If a project asks for information you did not expect to provide, or cannot explain in writing how the material will be used, stop and review the terms rather than assuming it is standard for the category.