Pacing: how much air to leave

A transcript cut is precise; the danger is that it is too precise. Every cut becomes a jump cut and the video "pops". These values came from real feedback.

Cut from the transcript, never from volume

A volume-only silence cut (ffmpeg silencedetect on the whole file) gave ~800 choppy pieces on a 34-minute class. Cutting from the word timestamps gave 120 pieces the creator loved. Use the waveform only to refine edges.

Planner algorithm (in words)

1. Silence map per clip: below -43 dB for at least 0.25 s.
2. Pause candidates inside kept speech: gap between two words ≥ 0.25 s and ≥ 85% of it is silence on the waveform (an untranscribed sound is not a pause).
3. Shorten only pauses longer than max_pause, down to pause_keep. Split the leftover air 40% before / 60% after.
4. Word edges: start = word start − 0.08 s, end = word end + 0.12 s, refined to the nearest silence edge within 0.25 / 0.3 s so soft consonants aren't clipped; never reach into a word that was cut.
5. Scene edges: the first range of a clip gets 0.5 s before the first word, the last range 0.6 s after the last word (only spoken words limit this; a lighter click or ambience may stay). Override per range in the keep-list (lead_in, tail). Look at frames first: stop before the creator reaches for the camera.
6. Snap to frames (start floors, end ceils, end exclusive), remove overlaps, merge touching pieces.
7. Write a review file: the transcript with cut words struck through + a timeline list with times.

Settings that worked