Remove Silence From Audio Without Chopping Breaths

To remove silence from audio you set three things: a threshold (how quiet counts as silence), a minimum duration (how long it has to last to count) and what stays (nothing, or a short pause). The third decides whether the result sounds chopped. Delete every gap completely and one sentence runs into the next with no air. Audacity's Truncate Silence, ffmpeg's silenceremove filter and edytlab's silence_finder and truncate_silence all use these settings, and they differ on the third, which is what this post is about.

The values below (-45 dB, 0.7 s, 0.3 s) are illustrations, not rules: breaths and room noise differ from one recording to the next, so find the silences and listen to them before you cut. The edytlab part needs edytlab, whose installers are unsigned, so macOS and Windows warn on first launch (Getting Started).

Threshold, minimum duration and what stays

• Threshold: the level below which audio counts as silence, in dBFS. Too low and background noise is never silent, so nothing is found. Too high and quiet speech, breaths and the ends of words count as silence and are cut. Audacity's Truncate Silence takes -20 to -80 dB, ffmpeg calls it stop_threshold, and edytlab's tools call it threshold_db.
• Minimum duration: how long the audio must stay under the threshold. Audacity calls it Duration (as short as 0.001 s), ffmpeg stop_duration, and edytlab min_silence_ms, which is 500 ms if you leave it out.
• What stays: Audacity's "Truncate to" keeps a length you choose, and "Compress to" keeps a percentage of the excess. ffmpeg's stop_silence keeps up to that long. edytlab's truncate_silence has no such setting and removes the whole silent stretch.

Truncate, cut or split: what is the difference?

Truncate shortens each silence and leaves some of it. Cut removes the silent stretch altogether. Split divides the audio at the silence into separate pieces and removes nothing. Audacity's manual draws the same line. Truncate Silence makes silences shorter by "deleting a section from the middle of the silent region", while Silence Finder, which Label Sounds has since superseded, is for when you "just want to split the recording into tracks at a specific point without removing the silences".

Truncate silence in Audacity

Audacity's manual describes Truncate Silence as reducing the length of passages whose volume stays below the threshold for at least the duration, by deleting a section from the middle of each, so some silence stays at both ends. These steps follow the manual and were not run for this post.

1. Select the audio and open Effect > Special > Truncate Silence.
2. Set Threshold (-20 to -80 dB) and Duration, the shortest silence to act on.
3. Choose the action. "Truncate Detected Silence" shortens every silence to the "Truncate to" length. "Compress Excess Silence" keeps the Duration and shrinks what is beyond it to the percentage you set, so long pauses stay longer than short ones.
4. Click OK.

Two limits from the manual. Truncate Silence only removes audio and does not reduce noise in the silences it keeps. And on audio with fades it can remove the quietest part of a fade, so apply it before adding fades.

Remove silence from audio with ffmpeg

FFmpeg's documentation says silenceremove removes silence from the beginning, middle or end of the audio. For the middle, a negative stop_periods makes it start over after each silence, and its examples use -1 to trim silence from beginning to end. stop_threshold is the level. stop_duration is the "duration of silence that must exist before audio is not copied any more". stop_silence is the "max duration of silence at end that will be kept after trimming". Detection defaults to the RMS over a 20 ms window, so a very short blip is averaged over that window and not judged sample by sample.

ffmpeg -i in.wav -af "silenceremove=stop_periods=-1:stop_threshold=-45dB:stop_duration=0.1:stop_silence=0.2" out.wav

Measured with silencedetect on a synthetic 11.25-second, voice-like file with pauses of 0.93, 1.42, 0.60 and 0.92 s, ffmpeg 6.1.1 left pauses of 0.29, 0.29, 0.32 and 0.29 s. The pause left in a long silence came out as about stop_duration plus stop_silence (0.1 + 0.2), not stop_silence alone: with stop_duration=0.7 and stop_silence=0.25, the 0.93 s pause stayed 0.93 s. The docs' own example reads as keeping at most stop_silence, so check what your version leaves by running silencedetect on the output.

The threshold matters more than the rest. With stop_threshold=-30dB instead of -45, the same file came out about a second shorter, because the synthetic breaths in it (noise bursts about 24 dB under the voice) then counted as silence. Real breaths differ, so measure before you cut. To list silences without changing anything, silencedetect prints each one's start, end and duration:

ffmpeg -i in.wav -af silencedetect=noise=-45dB:d=0.7 -f null -

Remove silence from audio in edytlab

edytlab has a finder and a cutter that use one rule. silence_finder is read-only and returns the silent ranges as start_sec and end_sec. truncate_silence removes them. Both take threshold_db and min_silence_ms. Both judge each sample frame on its own: a frame is silent when its loudest channel is under the threshold, and a run counts when it lasts at least the minimum. A single sample above the threshold ends the run, with no averaging window as in ffmpeg's default. So find first, which changes nothing:

On the synthetic file above, calling the tool directly returned three ranges: 2.42 to 3.35 s, 5.03 to 6.45 s and 10.33 to 11.25 s. At -30 dB the first range started at 2.09 s instead: a 0.35-second synthetic breath, peaking near -32 dBFS, had become part of the silence. That is the whole breath problem in one number, and it is why the finder comes before the cutter.

Remove the silences completely

truncate_silence deletes each qualifying range whole. It has no setting for a pause to keep, and it joins the two sides with no fade. On the synthetic file it took the length from 11.25 s to 7.975 s, which is exactly the three ranges; the 0.6 s pause, under the 700 ms minimum, was left alone. Labels placed on the track follow the audio.

Shorten the silences and keep a pause

The assistant has what it needs in silence_finder and cut_range, which removes a range of samples and shifts the rest left. For each silence it keeps 0.15 s at each end: it cuts from the start plus 0.15 s to the end minus 0.15 s, turning seconds into samples with the sample rate. It works from the last silence to the first because each cut moves everything after it, which is how the code of truncate_silence orders its own cuts. That sentence has not been run through a model here. Calling silence_finder and cut_range directly with those numbers on the synthetic file left pauses of 0.30 s in all three places, and the 0.6 s pause untouched.

This is the "truncate" of Audacity's manual, assembled from two tools. Breaths are not in the ranges when the threshold is under them, so they stay. Podcast production with an AI workflow puts it in the context of a whole episode.

No demo video removes silence. To see how a typed sentence becomes tool calls, and what the assistant reports back, the launch video follows a DJ job from request to export.

Split audio by silence

To split rather than remove, Audacity's route is Analyze > Label Sounds, which labels the sounds or the silences between them. Its threshold defaults to -30 dB, measured as peak, average or RMS level in 10 ms chunks, and a silence has to last 1 s by default. Then File > Export Audio with Multiple Files writes one file per label. FFmpeg's silencedetect gives you the times to cut at, and you cut the file yourself. In edytlab, silence_finder gives the ranges and split_clip splits a clip at a time you give, while label marks regions and export_labels writes them as Audacity-format label text. There is no step that exports each piece as its own file: export_multiple writes whole tracks as WAV files, one per track.

Questions about removing silence

How do I remove silence from audio for free?

Audacity's home page calls it a free, open-source audio editor, and FFmpeg's legal page says FFmpeg is licensed under the LGPL. edytlab is MIT-licensed; the only cost is the AI provider's tokens, unless you run a local Ollama model.

What does truncate silence mean?

Shortening silent stretches instead of deleting them. Audacity's effect of that name has a "Truncate to" length. edytlab's truncate_silence has the same name but removes the stretch completely, which is the cut in this post's terms.

Will removing silence cut my breaths?

It can, if the threshold is above the breath's level. Find the silences first with silence_finder, silencedetect or Audacity's Label Sounds, look at what falls inside them, and lower the threshold until the breaths are outside.

Does edytlab upload my audio to find silence?

No. The detection is local Rust. Your chat and the tools' results go to the AI provider you chose, and your audio does not; the local-first page has the table, and the tools reference lists every tool.

Sources

Outside pages were read on 2026-10-22. The ffmpeg measurements and the edytlab tool calls were run once, on synthetic audio generated for this post, with ffmpeg 6.1.1 and the edytlab tools called directly.

• FFmpeg Filters Documentation: silenceremove, read 2026-10-22
• FFmpeg Filters Documentation: silencedetect, read 2026-10-22
• FFmpeg: Legal, read 2026-10-22
• Audacity Manual: Truncate Silence, read 2026-10-22
• Audacity Manual: Label Sounds, read 2026-10-22
• Audacity Manual: Silence Finder, read 2026-10-22
• Audacity Manual: Exporting multiple audio files, read 2026-10-22
• Audacity: home page, read 2026-10-22
• edytlab tool reference (silence_finder, truncate_silence, cut_range, split_clip, label, export_labels, export_multiple), at commit b1678feb, read 2026-10-22
• edytlab source: truncate_silence.rs, at commit b1678feb, read 2026-10-22
• edytlab source: silence_finder.rs, at commit b1678feb, read 2026-10-22

edytlab is an open-source, local-first AI audio editor. Download the latest release or star it on GitHub.