How human does your synthetic person behave?

We score video of generated people with the same measuring we use on real ones, against the range of real human behavior, on six axes.

The six axes

• Pauses. How a person pauses while they talk.
• Blinks. How a person blinks while they talk.
• Gesture with speech. How the hands move with what is said.
• Voice. How the voice moves while a person talks.
• Expression timing. How the face moves with what is said.
• Gaze. Where the eyes go while a person talks.

How a system is scored

This is the method as it stands. Every score names the version it was made with.

Per axis: where the system sits among real people. Outside the human range, the score says which way it is off: too regular, too flat, too even or off the beat.

Together: whether real people show the same combination. Each axis can pass alone while the whole does not.

The headline: on how many of the six axes the system is inside the human range. It also names the first thing that gives the system away. There is no single “percent human” number.

From clip to clip: the score says how steady the system is across its clips.

The human range

It is the range of real behavior on each axis, measured on real people talking to the camera. A private score names the range it was read against. Submitted video is kept out of it.

What to send

• Several clips of one generated person. Each about ten minutes of speaking. Each clip is scored on its own.
• One person, talking to the camera. Alone on screen, facing the lens, speaking English.
• One take, as the system made it. No cuts, no editing and no music added afterwards.

Each clip goes through the same checks as a person’s video. The API docs list them.

Leaderboard

There are no scores yet. The leaderboard appears here when there are real ones.