Skip to main content
PixelTools
All articles
Guides5 min readby the PixelTools team

Remove Silence Does Nothing on Noisy Audio

Silence remover changed nothing? The pauses in your recording sit above the threshold. We measured how noise level and peak detection decide it.

Why the silence remover changed nothing

If the result is the same length as your original, the pauses in your recording are louder than the threshold, so none of them count as silence. Background hiss, a fan or room tone usually sits around -45 to -35 dB, and the default threshold of -40 dB sits right in that range. The tool measures the loudest single moment in every stretch, not the average, so a pause that sounds quiet can still fail. The fix is to raise the threshold toward -30 dB in steps of about 5 dB until the pauses count as silence, then check that quiet words survive. Our [remove silence tool](/remove-silence-from-audio) shows how many seconds it removed, so a result of zero tells you straight away that the threshold is too low for this recording.

Why a quiet pause still fails: the tool listens to peaks

The filter behind the tool, ffmpeg's silenceremove, runs here with peak detection. A stretch only counts as silent if its loudest sample stays under the threshold. Steady noise has peaks well above its average level. We generated white noise and measured it: noise with an average level of -44.8 dB peaked at exactly -40.0 dB, and noise averaging -35.2 dB peaked at -30.5 dB. So a hiss that looks like -45 dB on a meter touches -40 dB and is never treated as silence at the default setting. Cleaner noise averaging -55 dB peaked at -50.5 dB and was removed as expected. The practical rule: set the threshold a few decibels above the peak of your pauses, not above their average.

What we measured at each noise level

We built a 6 second clip with a 2 second tone followed by 4 seconds of noise, and ran the same filter the tool uses (native ffmpeg 6.1.1, 0.5 second minimum gap). At a threshold of -40 dB, noise peaking at -59.9 dB and -50.5 dB was cut, leaving 2.52 seconds. Noise peaking at -40.0 dB and -30.5 dB was untouched: the output stayed 6.00 seconds. Raising the threshold to -30 dB cut both of those too. These are synthetic signals, so real rooms will vary, but the pattern holds: the threshold either clears the noise peaks and removes the whole pause, or it does not and removes nothing. There is no partial removal for a pause that is slightly too loud.

What it does when it works: pauses shrink, they do not vanish

A working run does not delete pauses completely. Every pause longer than the shortest-gap setting is cut down to about that setting. With the 0.5 second default, pauses of 0.7, 1, 2 and 4 seconds all came out as roughly 0.52 seconds of silence in our test, and a 0.4 second pause was left alone. So a 4 second pause becomes half a second, which keeps speech sounding natural. If you want tighter pacing, lower the shortest gap toward 0.2 seconds, which in the same test cut a 10.2 second clip to 8.66 seconds. If you set the gap longer than your pauses, nothing is removed either: at 1.0 second, a clip with 1.0 and 0.8 second pauses stayed at 10.2 seconds.

How to fix it in a few steps

Open the tool and drop in the file. Run it once at the default and read the line that says how much silence was removed. If it says none, raise the silence threshold by 5 dB and run again, repeating up to about -30 dB. If speech starts to lose soft word endings, go back one step. Next, check the shortest gap: a long setting skips short pauses, and 0.5 seconds is a good default for speech. For very noisy files, the cleaner route is to reduce the noise first in an audio editor, then remove silence, because a threshold high enough to beat loud hiss will also eat quiet speech. Everything runs in your browser, so you can retry as often as you like without uploading the file.

The other failure: threshold too high

Raising the threshold has a cost. Anything quieter than the setting counts as silence, so soft consonants, breaths and trailing word endings can be cut when the pause that follows them is long enough. In our fade test, a 2 second fade-out from full level down to -48 dB lost 0.57 seconds at the end at -40 dB, and nothing at -60 dB. Music with quiet intros and fades, whispered speech and classical recordings are the usual victims. If the result sounds clipped, lower the threshold by 5 dB. For music, keep the threshold low (-50 dB or below) and use the tool only for dead air at the start or end of a track.

Frequently asked questions