Entry #6: Learning Faces and Voices

Omocube can now remember people by face and by voice. This is what lets Omocube address someone by name, keep an accurate sense of who said or did what, and keep what it knows about one person from bleeding into what it knows about another.

Omocube's camera view of an office with a person standing near a desk, boxed and labeled with their name at 49% confidence. Below it, a caption describing the scene and a transcript of the person asking 'Can you see me?' — Live camera view with person-aware caption and transcript based on face and voice recognition.

Recognition runs entirely on Omocube. Voice samples, face registers, and the names attached to them are never uploaded or matched against some larger database elsewhere. They get converted to numerical fingerprints that stay on-device.

1 — Enroll once

Face

A handful of shots at different angles and lighting

Voice

A few short recordings of them speaking naturally

2 — A fingerprint

Kept as a set of numbers used for comparison. Does not use raw photos or audio, and never leaves Omocube.

3 — Every time it sees or hears someone

Match found

Maya

Addressed by name.

No match

someone

Treated as “someone”.