The Agentic Data Company designs and collects audio datasets for the labs that train speech models. Its premise is that current speech models are trained largely on audiobooks, podcasts, scraped video and synthetic dialogue, none of which is real conversation, and that models trained this way miss intent and mishandle turn taking. The company records real people in real conversations to close that gap.
The vendor states that every hour of audio is consented, licensed and paid for at the source, and that datasets are built for frontier model training across languages and the long tail of how people actually talk. It works with the teams training these models, and invites them to describe where a model breaks so that a collection can be designed around the failure. The product is therefore a custom data service rather than a self serve tool. The founders, Christian Vestergaard and Matias Drejer, sign the company's statement of purpose.
The site also promotes Open Yap 1K, a publicly available conversational dataset.
Pricing is not published. Prospective customers use a request form to ask for a custom dataset.
