Where AI voice data comes from

Three sources of voice data

Open datasets. Free collections made for research. Many come from volunteers reading prompts, or from public domain audiobooks: five of the most widely used open voice datasets are built from the same LibriVox audiobook recordings. Flaunt Audio has compared the licences of 20 of them in its guide to open voice datasets for AI.

Scraped audio. Speech taken from videos, podcasts and websites. The people speaking were never asked, and the copyright usually stays with whoever made the original recording.

Commissioned recordings. Companies pay people to record speech to a specification, and every speaker agrees in writing before recording. This is the work Voices4AI castings are for.

Why companies still pay for new recordings

Free data runs out exactly where products fail: regional accents, telephone-quality speech, natural conversation and the words real customers use. It also has a gap no download can fix. An open licence covers copyright, not whether the people speaking agreed to their voices training a commercial product.

That is why the current castings ask for specific voices, such as Southern US, California, Mexican and mainland Spain accents and native Canadian French speakers.

What it means for you

Commissioned work is recorded for a defined project under a written agreement, so you know what you are agreeing to before you speak. The licence scope decides how your voice can be used, so read it. Our guides to selling your voice to AI and consent-cleared voice cloning explain what to look for.

A casting is one two-hour remote session using your own equipment, and a basic microphone is fine. No professional experience is needed. See what is open on the AI voice jobs board.