auto-editor: cut silence and pauses from talking-head videos by volume

Use the open-source command-line tool auto-editor (Unlicense, public domain) to make the first cut on video and audio. It decides which parts have content by volume (it can also use picture changes, black frames, or subtitle text), cuts out the rest, and outputs a trimmed file or an editor project. The repo ships with five official Skills. This page is the version written for agents. For the step-by-step guide for people, see Let an AI Agent Cut Every Dead Pause Out of Your Talking-Head Video.

Triggers

• The user wants to cut silence, pauses, or empty stretches at the start and end of a talking-head video, screen recording, or podcast.

• The user wants to shorten a recording, or fast-forward quiet parts instead of cutting them.

• The user wants to import the auto-cut result into Premiere, DaVinci Resolve, Final Cut Pro, Shotcut, or Kdenlive to keep editing.

• The user wants a transcript or subtitles from a video, or wants to cut by what was said.

Hard rules

• Never overwrite the original. Output always goes to a new file (default name <original name>_ALTERED.<extension>, or set with -o).

• Preview before rendering. For long videos, run --preview first, report to the user in plain language how much will be cut and how long the result will be, and render only after they confirm.

• Ask before installing. The official binary is not signed, so the system may show an "unidentified developer" warning. Explain why, and let the user decide whether to continue.

• Don't install with pip. The official project no longer publishes the command-line version to pip. The pip and apt versions are old.

• Don't promise it can detect slips of the tongue. It works on signals and cannot understand content. Repeated lines and slips must go to the user through a transcript for confirmation.

Install

Check first, then use the official recommended method:

auto-editor --help >/dev/null 2>&1 && echo "already installed" || echo "missing"
# macOS (Homebrew)
brew install auto-editor
# Other systems: download the binary for your platform from GitHub Releases,
# rename it to auto-editor, run chmod +x ./auto-editor on macOS/Linux, then put it in a PATH directory
# Arch Linux: yay -S auto-editor

Official Skills (five in total: cutting, effects, export, transitions, and transcription):

npx skills add WyattBlue/auto-editor

Optional: if yt-dlp is installed, you can pass a video URL directly as the input.

Pick a command by need

NeedCommandCut silence with default settingsauto-editor video.mp4Set the output fileauto-editor video.mp4 -o out.mp4See what would be cut, without renderingauto-editor video.mp4 --previewShow file infoauto-editor info video.mp4Adjust the volume line (lower it if quiet speech is being cut)--edit audio:threshold=0.03 or --edit audio:-30dBBuffer before and after each sentence (0.2 seconds each by default)--margin 0.3s,1.5sec (0.3 seconds before, 1.5 seconds after)Cut by picture changes (silent screen recordings)--edit motion:threshold=0.02Keep a segment if either sound or picture has activity--edit "(or audio:0.03 motion:0.06)"Fast-forward silent parts instead of cutting them-w:0 speed:8Force keep or force cut a segment--keep 0,30sec, --cut -10sec,endReview what was cut (inverse view)--when-active cut --when-inactive nilExport an editor project--export premiere, resolve, final-cut-pro, shotcut, kdenliveReduce file size-crf 23 -preset medium or -b:v 230kNormalize loudness-anorm ebuTranscript or subtitlesauto-editor whisper video.mp4 <model file> --format srt -o out.srt

Workflow

1. Run auto-editor info to check the duration and audio tracks. If there are several tracks (for example, mic and background audio recorded separately), ask the user which track to judge by.

2. Run --preview with default settings and report the original length, the length after cutting, the number of cuts, and the longest silence. If more than half would be cut, suspect the volume line is too high first.

3. After the user confirms the settings, render. For talking-head videos, default to a "less before, more after" buffer such as --margin 0.2s,0.4s to avoid clipping the ends of sentences.

4. If the user will keep editing by hand, export the matching editor project file at the same time.

5. If the user wants slips and repeats found: use auto-editor whisper to create a timestamped transcript. whisper.cpp model files must be downloaded separately. On macOS 26 or newer, you can use apple to call the system recognizer. List the likely repeated sentences with timestamps, and wait for the user's confirmation before cutting.

Limits

• It works only on signals (volume, picture changes, black frames, subtitle text) and does not understand meaning. If someone misspeaks and then repeats the line loudly, both takes are kept.

• With heavy background music or ambient noise, quiet parts still have sound, so cutting fails. Record clean voice and add music later.

• In screen recordings, silent action sections are cut as silence. Switch to motion or --keep for these.

• Transcription needs a speech model download, and the first run is slow.

Output format

After cutting, report to the user:

Original length: mm:ss
Length after cutting: mm:ss (shortened by xx%)
Settings used: volume line / buffer / other
Output file: <path to edited file>
Project file: <path, or "not exported">
Places to review by hand: <list of timestamps, or "none">

In one sentence

auto-editor removes the parts with no sound. Which of the remaining parts to delete is for the user to decide from the transcript.