Silence cutting compared: auto-editor against FFmpeg silencedetect on the same clips
auto-editor and FFmpeg's silencedetect both find quiet stretches by loudness. We ran both on the same three 60-second clips at matching levels and compared the cuts, the speech cut with them and the run time.
Tested by us on Nov 6, 2026, on the clips and machine listed below. Every number on this page was measured in that run.
The result
On our three film clips the two tools mostly agree, and the difference that matters is not the cuts. At the stricter level (8% for auto-editor, -22 dB for FFmpeg) clip tos_b lost 5.42 s to auto-editor in 11 cuts and 5.48 s to FFmpeg in 7 cuts, and FFmpeg took more of the speech with it (2.21 s against 1.71 s by the film's subtitle times). At the default-matching level neither tool cut anything on tos_b or tos_c.
What differs is the work around the cut. FFmpeg's silencedetect finds the silences in about a quarter of a second on a 60-second clip, but it only reports them; you then write the filter that removes them. auto-editor finds the cuts and writes the video in one command, in 26 to 38 seconds here. Our own FFmpeg cut pass, a second command that re-encodes the video, took 24 to 25 seconds (one run each).
What failed or surprised us
- The film has a music score under its dialogue, so there is little true silence. Three of the six FFmpeg runs found nothing to cut, and so did auto-editor at its default. This comparison shows how the two tools behave on music and dialogue, not on a clean voice recording, where both would cut far more.
- The tools do not line up exactly. At -22 dB FFmpeg found seven silences on tos_b where auto-editor made eleven cuts, because auto-editor pads each side of a cut with a margin and merges or drops tiny ones by its own rules, while silencedetect reports a silence only when it lasts at least 0.5 s. We matched the loudness levels by arithmetic (4% of full scale is -28 dB, 8% is -22 dB), not by a calibration, so treat the pairing as approximate.
- FFmpeg took more speech on tos_a and tos_b than auto-editor did. We measure speech cut against the film's subtitle times, which are approximate, so a difference of a few tenths of a second is within that error.
- FFmpeg's silencedetect run was too short (0.23 s) for our memory sampler to catch, so we show no memory figure for it. auto-editor's timings come from our earlier auto-editor test, not from this session.
The cut maps
Each bar is one 60-second clip. Dark parts are kept, light parts are removed. Where a bar is all dark, that method cut nothing.
tos_a
tos_b
tos_c
What we tested and how
- Tool: auto-editor and FFmpeg silencedetect version auto-editor 29.3.1 and ffmpeg version 8.1.2-full_build-www.gyan.dev, Unlicense (auto-editor), LGPL/GPL (FFmpeg build).
- Settings compared: auto-editor threshold 4% (default) and 8%; FFmpeg silencedetect noise -28 dB and -22 dB, minimum silence 0.5 s.
- Input clips: Tears of Steel, (CC) Blender Foundation, mango.blender.org, CC-BY 3.0. 60 s excerpts re-encoded to H.264 720p.
tos_a: 60 s starting at 20 s of the film, SHA-256945262dcbe106f1e...tos_b: 60 s starting at 180 s of the film, SHA-2564097773b615cddfa...tos_c: 60 s starting at 360 s of the film, SHA-2563d3a8cd015310349...
- Machine: AMD Ryzen 7 7735U (8 cores, 16 threads), integrated Radeon 680M, not used (CPU-only runs), 31.2 GB RAM, Windows 11 10.0.26300. The laptop was shared with other work, so timings are what a busy machine gives, not a clean benchmark.
- Runs: one cold run, then three timed runs per clip and setting. Run on 2026-11-06.
The same three 60-second clips of Tears of Steel, two loudness levels per tool. auto-editor: its default threshold (4% of full scale) and a stricter 8%. FFmpeg: silencedetect at the same two levels written in decibels (-28 dB and -22 dB), with a minimum silence of 0.5 s. For FFmpeg we then removed the silences with select and aselect filters and measured the real length of the file that came out. The auto-editor numbers are read from the result files of our auto-editor test. The FFmpeg silencedetect pass was run once cold and three times timed per clip and level.
Speech cut is measured against the cue times of the film's official English subtitles, the same measure as the auto-editor test.
Measurements by clip
Every method on every clip
| Clip | Method and setting | Cuts | Removed (s) | Subtitled speech cut (s) | Output length (s) | Median run (s) | Peak memory (MB) |
|---|---|---|---|---|---|---|---|
| tos_a | auto-editor, threshold 4% | 1 | 1.33 | 0.46 of 25 | 58.67 | 26.4 (finds the cuts and writes the video) | 490 |
| tos_a | FFmpeg silencedetect, noise -28dB, min silence 0.5 s | 1 | 1.7 | 0.7 of 25 | 58.31 | 0.23 to find the silences, then 23.7 to write the video (one run) | not measured (run too short) |
| tos_a | auto-editor, threshold 8% | 3 | 2.08 | 0.58 of 25 | 57.92 | 23.6 (finds the cuts and writes the video) | 487 |
| tos_a | FFmpeg silencedetect, noise -22dB, min silence 0.5 s | 1 | 1.83 | 0.83 of 25 | 58.17 | 0.23 to find the silences, then 20.5 to write the video (one run) | not measured (run too short) |
| tos_b | auto-editor, threshold 4% | 0 | 0 | 0 of 27.2 | 60 | 38.2 (finds the cuts and writes the video) | 493 |
| tos_b | FFmpeg silencedetect, noise -28dB, min silence 0.5 s | 0 | 0 | 0 of 27.2 | 60 | 0.23 to find the silences (nothing to cut) | not measured (run too short) |
| tos_b | auto-editor, threshold 8% | 11 | 5.42 | 1.71 of 27.2 | 54.58 | 33.7 (finds the cuts and writes the video) | 487 |
| tos_b | FFmpeg silencedetect, noise -22dB, min silence 0.5 s | 7 | 5.48 | 2.21 of 27.2 | 54.54 | 0.23 to find the silences, then 24.8 to write the video (one run) | not measured (run too short) |
| tos_c | auto-editor, threshold 4% | 0 | 0 | 0 of 24.5 | 60 | 26.1 (finds the cuts and writes the video) | 482 |
| tos_c | FFmpeg silencedetect, noise -28dB, min silence 0.5 s | 0 | 0 | 0 of 24.5 | 60 | 0.23 to find the silences (nothing to cut) | not measured (run too short) |
| tos_c | auto-editor, threshold 8% | 2 | 0.71 | 0 of 24.5 | 59.29 | 34.8 (finds the cuts and writes the video) | 482 |
| tos_c | FFmpeg silencedetect, noise -22dB, min silence 0.5 s | 0 | 0 | 0 of 24.5 | 60 | 0.23 to find the silences (nothing to cut) | not measured (run too short) |
Reproduce it
# auto-editor, default threshold (4%) and the stricter one (8%)
auto-editor clip.mp4 -o out.mp4
auto-editor clip.mp4 --edit audio:threshold=8% -o out.mp4
# FFmpeg: find the silences (min length 0.5 s) at the same two levels, 4% = -28 dB and 8% = -22 dB
ffmpeg -i clip.mp4 -vn -af silencedetect=noise=-28dB:d=0.5 -f null -
ffmpeg -i clip.mp4 -vn -af silencedetect=noise=-22dB:d=0.5 -f null -
# FFmpeg: cut them out (kept ranges from the silencedetect lines, written as between() terms)
ffmpeg -i clip.mp4 -vf "select='between(t,0,10.2)+between(t,12.4,60)',setpts=N/FRAME_RATE/TB" \
-af "aselect='between(t,0,10.2)+between(t,12.4,60)',asetpts=N/SR/TB" -c:v libx264 -crf 20 -c:a aac out.mp4The commands for both tools. In the last pair, replace the between() terms with the kept ranges that silencedetect printed, which are the gaps between its silence_end and the next silence_start.
Method and limits
- Three clips of one film with music under the dialogue. A clean voice recording would give very different cut counts.
- Two settings per tool, matched by arithmetic. Neither tool was tuned.
- Run times are from a laptop shared with other work; the CPU load before each FFmpeg run is in the data file.
- We did not listen to the cuts. The speech measure counts overlap with subtitle times and says nothing about whether a cut sounds natural.
- Footage: Tears of Steel, (CC) Blender Foundation, mango.blender.org, CC-BY 3.0.
Page checked on Nov 4, 2026.