The Multimodal Wave: Voice, Video, and Image AI Training Work

2026-09-19 · Expert Match AI team

For two years the AI training market was a text market: rate this answer, grade that proof, rewrite this paragraph. That is changing on our board in plain sight. Of the 316 roles we track today, 37 are about capturing or labelling something other than text - audio, video, images, faces, voices. Eight weeks ago that category was a handful of listings. Today it is the fastest-growing shape of work at the entry end of the market, and it has a pay structure all its own.

This post is what the multimodal tier looks like right now, from the live listings.

What the labs are collecting

Three kinds of work make up almost all of it.

Audio transcription and transcript editing. Alignerr has posted a wave of paired roles - transcriptionist plus transcript editor - in Brazilian Portuguese, French, German, Hindi, Italian, Simplified Chinese and Spanish (Spain), all at $10–35/hr. micro1 has its own set in Cantonese ($20–35), Mandarin ($20–30), Dutch ($25–30), German ($10–25), Tamil and Telugu ($10–24). Braintrust lists a Voice Recording Contributor at $20–40. That is fourteen-plus language-specific audio roles live at once; the labs are building speech models beyond English and need native ears to produce and correct the ground truth.

Video capture. micro1 is recruiting people to be the camera: a 360° Video Recorder (Insta360 / GoPro MAX) at $20, an Exocentric Video Recorder at $20, a Vlogger role for the same equipment at $20, a Smartphone Video Recorder at $14–15, a US-based Data-Video Generalist at $13–15, and a Portuguese-language smartphone video recorder at $6. The listings describe continuous movement through changing environments, vehicle navigation, even animal-mounted footage. This is training data for models that need to understand the physical world - robotics and embodied AI - and it needs thousands of hours of ordinary first-person footage.

Faces, images and sound annotation. The outlier is micro1's AI Facial Data Collection Associate at $139–140/hr, with sibling and twin variants (the twins role has no listed rate). That rate is not a typo on their side; it is a scarcity price for a very specific dataset, and it is the one multimodal listing that pays like the expert tier. Below it: a Music & Sound Annotation Expert at $30–45, a Video Annotation Specialist at $8–10, and a set of video editor roles (Premiere, Final Cut, generic) at $15–80 that sit closer to conventional freelance work than to training data.

Alignerr's Central American Spanish healthcare call reviewer and QA labeler roles ($15–20) are audio too - reviewing recorded calls for a healthcare AI - and show the pattern moving into regulated domains.

What it pays against text

Put the multimodal roles next to the text-based board and the picture is clear:

Multimodal collection sits squarely in the annotation tier, and mostly at the bottom of it. The reason is structural: the listing needs you for what you have (a camera, a voice, a language) rather than what you know, and that is a large pool. The two exceptions are both scarcity plays - the facial-data role, and to a lesser degree the rarer-language transcription slots, where the rare-language premium has mostly collapsed but the queues are real.

Expect these rates to behave like the annotation tier in practice too: per-task or per-hour-of-footage pay quoted as hourly, and batches that close when the dataset is full. Run the effective-rate arithmetic before treating a $20/hr capture role as $20/hr.

Who it actually suits

Native speakers of the listed languages who want a fast start. Transcription is the one multimodal category where the requirement - fluency - is real and non-trivial, which thins the pool. If you speak Hindi, Cantonese or Brazilian Portuguese natively and have no professional credential, an audio queue is a more realistic first platform than a generalist text queue, and it plugs directly into the first-90-days plan as an anchor.

People who already own the equipment. The 360° and action-camera roles require you to have an Insta360, GoPro MAX or DJI Osmo. If you do, $20/hr to walk around with it is real money for something you were doing anyway. If you do not, the equipment cost eats the first several weeks of pay.

Not credentialed professionals. A physician or an attorney has no business in a $15/hr transcription queue when Surge, micro1 and Mercor list physician tiers at $250–450 and attorney tiers at $140–1,000. Every hour of multimodal collection is an hour not spent getting into a standing expert pool.

What to watch

Multimodal is where the board is growing at the bottom, and it is also where listings churn fastest - Alignerr's transcription wave arrived over about three weeks and, if the pattern from previous batch-shaped waves holds, will thin out as quickly once the datasets fill. Apply early or not at all. The jobs board sorted by newest shows each week's arrivals, and the salary report tracks whether the ML, Data & AI category, where most of these roles land, moves off its current $63/hr average as the mix shifts.

Published by the Expert Match AI team. Role counts and rates are from our tracker as of September 15, 2026 (316 roles tracked, 222 with disclosed hourly pay, 10 platforms); the multimodal count is a title-and-description match for audio, video, image, voice and transcription work and is approximate. All figures are listed rates, not accepted offers. Some outbound application links on this site carry disclosed referral codes; rankings and recommendations are never influenced by referral payout.