AI needs to
meet more people.

Somewhere inside the brief for an intelligent car, there were twins. Also children, older adults, and pets. There were daytime recordings and nighttime recordings, ordinary cameras and infrared cameras. The assignment described in Dataocean AI’s September 2026 case study sounds, at moments, like a casting call for an unusually demanding family film. It was a training dataset.

The unnamed electric-vehicle customer wanted examples for systems that recognize occupants, interpret gestures, monitor passenger states, and handle speech. A car has to encounter people in their inconvenient variety. Before engineers can teach it what to notice, somebody has to arrange for that variety to show up.

A casting call for the machine

Dataocean AI translated the cockpit brief into quotas: participant groups, vehicle models, time of day, scenarios, and tasks. It organized recruitment pools and backups. Selected pilot videos went to early review so the customer could check actions, camera results, and formats before larger batches followed.

That last detail deserves attention. A thousand correctly labeled recordings of the wrong action remain the wrong recordings. In this project, unified task identifiers connected collection, annotation, and delivery. Noncompliant material was corrected or collected again. Quality control reached all the way back to the moment an example was made.

Then came the physical complications. Unreleased vehicles required confidentiality controls. Powered-on stationary cars required braking and wheel restraints. Moving vehicles required qualified drivers and speed limits. Here, AI development included managing car keys.

The recording room gets complicated

A separate case study describes approximately 2,000 hours of natural duplex conversation delivered in five months, involving more than 1,000 participants. Duplex conversation captures people responding, overlapping, and changing pace. A neat script would remove some of the behavior the customer needed.

Recruitment therefore screened for demographic and language requirements. Verification tools checked participant uniqueness and duplicate voices. Recording, annotation, and review ran in parallel, with staffing adjusted as workloads shifted. Dataocean AI says 300 to 500 annotation or quality specialists could be mobilized during peaks.

The resource bill becomes visible even without a dollar figure: five months, a large participant pool, specialist review, and controlled recording environments. Hardware clock synchronization kept microphone channels aligned within milliseconds. For a buyer, these are procurement questions worth asking alongside the price per hour.

Twenty years of listening

The business began in 2005 with speech-recognition and speech-synthesis data. Its official history dates expansion into computer vision and natural language processing to 2009, autonomous driving to 2017, and generative AI to 2022. Formerly Speechocean, it recorded a brand upgrade in 2024.

The commercial offer has two useful doors. Buyers can license existing speech, image, text, or multimodal datasets. Or they can commission a project with particular speakers, settings, sensors, annotations, and acceptance rules. The second route purchases an organized production effort as well as files.

Fit matters more than volume. A team with a narrow, well-covered use case might find an existing dataset sufficient. A team needing an unusual accent, a particular sensor, or a new interaction needs to examine the actual examples. Collection conditions and permission to use the material belong in that decision, along with the model’s performance on held-out tests.

Customers include enterprise AI teams, automotive developers, and researchers. The company lists Microsoft, NVIDIA, Amazon, Google, and Hyundai among its customer references. These references establish its intended market; the detailed production accounts concern unnamed clients.

Its DOTS platform supports configurable workflows, automated pre-labeling, assisted annotation, and quality inspection. The company advertises more than 200 annotation task types. The practical attraction is connecting stages that might otherwise become separate queues, spreadsheets, and opportunities for mismatch.

“Enhanced data quality results in more efficient models, which in turn accelerates deployment.”

Dataocean AI’s stated approach

When the laboratory runs out of world

The company’s embodied-AI case study makes the argument physical. A world-model customer needed examples from unstructured environments. Laboratory collection missed variation in lighting, occlusion, materials, and interaction paths. Separate collection and review processes also lacked shared operating standards.

The response was distributed collection in homes, offices, retail, and industrial settings, using first-person capture, UMI collection, and robot teleoperation demonstrations. Scheduling, alignment, annotation, and delivery were brought into a shared workflow. Dataocean AI reports capacity exceeding 50,000 hours of usable data per month in that program.

That is a supplier-reported capacity figure. The more transferable detail is the decision behind it: make the process repeatable before multiplying locations. A coordinated pipeline still depends on choosing environments that represent the task the robot must perform.

A dolphin you can actually try

There is also a public way to inspect the company’s technical work. Dolphin, developed with Tsinghua University, supports 40 Eastern languages and 22 Chinese dialects. Its repository describes training on more than 210,000 hours of proprietary and public audio. Released code and model weights use Apache 2.0.

The May 2026 Dolphin-CN-Dialect paper addresses an instructive problem: dialect data is unevenly distributed. The researchers changed sampling and tokenization and added streaming capabilities. Simply accumulating recordings would leave the balance problem intact. Engineers can download the released models and test recognition against their own speech conditions.

Dataocean AI competes with providers such as Shaip in speech collection and licensing, and Scale AI in annotation workflows. Its pitch combines speech experience, multilingual sourcing, and production across modalities. The sensible buying test is specific: request samples, define acceptance criteria, check usage rights, and compare results on the intended task. An impressive catalogue cannot compensate for examples that miss your users. Start with the people the model needs to meet.