AI

The clip finder: AI that finds the moments in hour-long live streams

Own project · 2 min read · Otto Boström

A live stream can be ten hours long. The moments worth cutting into short videos add up to perhaps three minutes. The clip finder is the workflow I built to find them without anyone having to watch it all.

The problem

Looking for clips by hand means someone watches the whole stream again afterwards. It rarely gets done, and then hours of material are left unused. The same was true of my archive of older videos.

The solution

• Everything that is said is written out with speech recognition that runs on my own computer.
• Three signals are weighed together: what is said, how the sound level moves and how active the chat is.
• An AI assistant goes through the candidates and writes one suggestion per clip, with the time, a title and a short reason.
• The result is a finished brief for the person who edits, one document per stream.

Results

• 36 archived streams were gone through in one night and became 32 finished briefs for the editor.
• 1,397 older videos were gone through and gave 5,337 ideas for short videos.
• Writing the brief for one stream takes about four minutes, with nobody sitting at the computer.

What I learned

AI makes mistakes in a predictable way. The speech recognition can, for example, repeat the same line when the sound is loud and messy. So the workflow is built so that a person always makes the final decision: the tool suggests, the editor chooses. I take the same approach in client work.

For your company

Swap the streams for recorded meetings, customer calls, long documents or a full inbox. The pattern is the same: let AI go through the pile, point out what is worth a person's time and leave the rest.