Auto reframe video to 9:16, framed on whoever's talking

For a speech video, reelaxe finds the faces in each shot. When everyone fits in the vertical frame together, it keeps them all in. When they don't, it has to pick one.

To pick, it looks at two things at once: how much each mouth is moving, and how loud the audio is at that moment. A mouth that moves while there's speech counts. A nod or a silent laugh in a quiet moment doesn't. Over every second it compares the faces, and one counts as talking only when it clearly beats the other.

The frame doesn't jump on every "mm-hm". It cuts to a new speaker only once they've been the one talking for a second and a half, and then it cuts to the moment they started. When both people talk at once, the frame stays where it is. Each new shot opens on whoever talks first.

A change of speaker is a cut, not a slow pan across the room. That's how an editor would do it.

Within a shot, the frame holds still. It only moves when the person moves well out of the frame and stays out. A small lean or a hand gesture doesn't move the camera. If someone walks across the shot, the frame follows them.

Face detection also holds on when a face is briefly lost, like a turn to profile or a hand in front of the mouth. A one-off miss won't hand the frame to the other person.

• Only one face, or faces that fit together: framed as is. No guessing needed.
• Nobody clearly talking: the frame holds on one person.
• The audio can't be read: it frames one person per shot and the job message says it ran "without speaker detection".
• No face in a shot: it looks for a person instead, then falls back to the centre of the frame.
• Music videos: speaker detection is for speech. For songs and rhymes it follows the main subject in each shot, like a person or an animal.

Every clip is vertical 9:16 at 1080×1920, ready for Reels, Shorts and TikTok. Square (1:1) and wide (16:9) are coming soon.