The pipeline
Four passes over your footage, none of them yours.
Keptbits runs the route an editor would, in the order an editor would. It just does it while you get coffee.
- 00:00
Bring the source
Drop the episode in. A two-hour interview uploads in the background, straight from your browser — the file never sits on a server of ours in between. Then narrow the range if you only want part of it read.
- 00:04
Read every word
A word-level transcript with speaker separation. Nothing is summarised: the scoring pass works from what was actually said, with timings accurate enough to cut on.
- 00:09
Judge the moments
Every candidate is scored on whether it opens and resolves inside itself. A clip that needs the ten minutes before it is worth nothing, so it does not survive.
- 00:12
Cut and caption
Boundaries snap to real word timings, the frame cuts to whoever is speaking rather than drifting, and captions burn in word by word in the style you picked.
The two decisions that matter
Where a clip starts and stops
The model proposes a window; the window is then snapped to real word timings, preferring the start of a sentence. That snapping does more for how finished a clip feels than the model’s own precision does — a cut that lands mid-syllable reads as broken no matter how good the moment was.
Who is in frame
The transcript already knows who is speaking, so the only question left is where they are. The frame cuts to them and holds — it never slides across the room, because a sliding crop reads as a mistake rather than an edit.
What it costs to run
One token is one minute of footage read. Reading is the part that costs money — the transcript and the scoring pass are both priced per minute of source. Cutting a clip from a transcript we already hold is close to free, which is why clips are unlimited and re-exports never cost you anything.
It also means narrowing the range genuinely saves you money. Pointing a two-hour recording at the twenty minutes that matter costs twenty tokens, not a hundred and twenty.