AI Video Editing Flows

AI talking head video editing in one pass

This workflow edits a talking-head video in one pass: remove pauses, tighten the cut, add captions and finish with a light polish. The page gives the acceptance criteria, the steps with prompts and commands you can copy, and a free starter download. It counts human finishing time against a manual edit.

This page is built but kept out of search until it carries the evidence it needs: Needs a recorded talking head we may use; in testing.

The result

We ran the one-pass edit with free command-line tools and no coding agent. It produced a captioned, loudness-levelled file from each of three 60-second clips in about a minute and a quarter on a laptop CPU, with the full measurements on We edited three talking clips in one pass: cut, word-timed captions, loudness, burn-in.

What we did not run: an agent doing this job from a brief. That needs an agent account, so this page describes a scripted pass and says nothing about how well an agent does it.

What failed or surprised us

  • Single-pass loudness missed its -16 LUFS target on two of three clips (finals at -16, -13.9 and -16.8). Two passes measured -15.9, -16 and -16.
  • One word got a 5.5 s timing span, which held a three-word caption on screen for 6.4 s.
  • The cut was small on film footage with music. We have not run it on a clean talking-head recording.

The steps

  1. Cut the quiet parts: auto-editor clip.mp4 --edit audio:threshold=8% -o cut.mp4
  2. Get word times from the cut clip with faster-whisper (small.en, word timestamps on).
  3. Group the words: at most three per caption, a new caption after a pause of half a second or a sentence end, each shown at least 0.35 s.
  4. Level and burn in: loudnorm I=-16 in two passes (measure, then apply the measured values with linear=true) and the subtitles filter in the Shorts style.

The full script and the two-pass commands are on the test page, under Reproduce it.

What it took on this laptop

A median of 68.4, 77.8 and 72.8 seconds for the three clips, with peak memory under 1 GB, on an AMD Ryzen 7 7735U with no GPU in use. The laptop was shared with other work.

Method and limits

  • Tested on three 60-second film clips, not on a recorded talking head.
  • A person still has to review the whole export; we looked at frames from each file.
  • Our method is on How we curate and test AI video workflows.

The workflow is run on real footage and judged against written acceptance criteria. We record the time to an accepted export, including the human finishing time, the tool versions, and every step that failed.

Page checked on Nov 4, 2026.