Talking-head video for avatar training, with likeness rights that name the use

A talking head dataset is mostly a consistency exercise. Framing stays fixed from mid-chest up, head pose is held within an agreed range, and lighting is diffuse and constant across the session, so the model learns faces rather than shadows. For lip movement, frame rate matters more than resolution: bilabial closures that smear at 25 fps resolve cleanly at 60.

Sync is the quiet requirement, because a lip-sync model trained on footage that drifts half a frame learns the drift as truth. Audio is locked to video at capture rather than corrected afterward. Scripts are phoneme balanced so every viseme appears often enough to learn from, and each session includes expression range, silences, and listening segments instead of a single neutral read.

Scripted reads give you controlled viseme coverage and clean alignment between text, audio, and lip movement, which is where most lip sync training data starts. Spontaneous monologue adds the hesitations, head movement, and natural prosody that scripted speech suppresses. Two-person dialogue adds turn-taking: interruptions, backchannels, and the gaze shifts that mark a speaker handing over the floor.

Dialogue also fills the gap most digital human datasets leave open: the listening face. An avatar in conversation spends half its time not speaking, and if the corpus contains only speech, the model has nothing honest to generate between turns. Recording both sides of every conversation produces training data for AI avatars that speak and avatars that listen, from the same session.

Face video sits next to biometric data in European privacy law, and an avatar is the most literal reuse of a likeness there is. Every participant signs a release that names the intended use in plain language: avatar generation, lip-sync training, synthetic media, whichever applies to your project. The consent chain is documented per participant and delivered with the data, so a rights question two years from now has a paper answer.

That is the line scraped footage cannot cross. Public talking face datasets built from broadcast or YouTube material are generally licensed for research only, and the people on screen never agreed to become a digital human. A custom collection costs more than a download, and it is the difference between a model you can ship and one you can only publish about.