How to cut silence in a video
A complete pass takes about ten minutes: two settings, one scan, and a review. This guide explains what the machine is measuring, so that when the first attempt is wrong you know which number to move.
What a detector actually measures
Silence detection is loudness measurement. The tool reads the audio track, computes a level for each short window of time, and compares it against a threshold. Any run of windows that stays below the threshold for longer than a minimum duration is marked as silence. What survives is stitched back together, and that stitched sequence is your edit.
Two consequences follow, and both explain most of the surprises. First, silence means quiet, not empty: a room tone, a fan, or a hum can keep a passage above the threshold even though nobody is speaking. Second, the detector cannot see the picture. A stretch where you held up an object or left a diagram on screen is silence to the audio and content to the viewer.
The three settings
Threshold, or sensitivity. This is the line between noise and speech. It has to sit above your noise floor and below your quietest speech, and how much room there is between those two is decided by your recording, not by the tool. A close microphone in a quiet room leaves a wide gap and is easy. A laptop microphone in a room with an air conditioner may leave almost none.
Minimum length. How long a quiet passage must be before it is worth removing. This is the setting people get wrong first, because it seems obvious that shorter is more thorough. It is not: ordinary speech is full of gaps of a few hundred milliseconds between words and clauses, and cutting those destroys the rhythm of the sentence. How long should a pause be works through where to start.
Margin. How much silence to keep on each side of the speech that survives. Cutting exactly where the level crosses the threshold clips the soft beginnings and endings of words, especially consonants like s, f, and t, which are quiet compared with vowels. A margin of a couple of tenths of a second is usually the difference between a join you hear and one you do not.
Doing the pass
- Open the file. In CutNoodle the browser reads it from disk, so there is no upload and no waiting proportional to file size.
- Set the minimum pause and the sensitivity. Around three seconds and a middle sensitivity is a sane first attempt for a normal room. You are aiming to be roughly right, not exactly right.
- Scan. Only the audio track is analysed, which is why this is quick even on a long recording.
- Read the result before watching it. The timeline shows the surviving clips against the waveform. If half the video vanished, or nothing did, you can tell at a glance and fix the setting before spending any time on playback.
- Review the joins. Play across each cut with sound on. This is where the pass earns its trust.
- Export. An MP4 with H.264 video and AAC audio, written locally.
Fixing what went wrong
Nothing was cut. Your noise floor is above the threshold. Raise the sensitivity. If raising it far enough starts eating speech, the recording itself is the problem: the gap between noise and voice is too narrow, and cleaning the audio first will do more than any setting.
Words are clipped. Lower the sensitivity so quiet speech is not read as silence, or widen the margin so more air is kept. Try the margin first: it fixes the common case without changing which passages are detected.
The result stutters. Raise the minimum length. If it still feels breathless at a generous setting, you are cutting pauses that carried meaning, and the fix is to put some back by hand rather than to keep adjusting numbers.
A specific cut is wrong. Do not chase it with settings. Merge the two clips on the timeline to restore the pause, or drag a clip edge to move the boundary. One manual correction is cheaper than a global change that breaks three other places.
Reviewing without watching it all again
You do not need to watch the whole video. You need to hear the joins, which are the only places the file changed. Jump between cut points, listen across each one, and move on. On a twenty-minute recording that is a few minutes of work and it catches essentially everything: clipped words, pauses that mattered, and cuts that landed while something was happening on screen.
The thumbnails on the video track help with that last category. A cut in the middle of a visual demonstration usually looks wrong before it sounds wrong, and merging the neighbouring clips puts the time back in one action.
Common questions
Why did nothing get cut?
The background noise in your recording is louder than the threshold, so no passage ever reads as silent. Raise the sensitivity until the quiet stretches register.
Why does the result stutter?
The minimum pause length is too short. Normal speech contains many gaps under a second, and removing them all produces a machine-gun rhythm. Raise the minimum until only real dead air qualifies.
Why are the ends of my words missing?
The cut is landing too close to the speech. Either the threshold treats quiet consonants as silence, or the margin kept around each phrase is too small. Lower the sensitivity or widen the margin.
CutNoodle is free and runs in your browser. Open the editor and drop a video in: nothing is installed and nothing is uploaded.