The gaps, the dead air, the three seconds before you started talking. We measure how quiet your recording's quiet actually is before deciding what counts as silence — because on a real microphone in a real room, the usual answer finds nothing.
Drag & drop audio or video, or click to browse
MP3, WAV, FLAC, M4A, OGG, MP4, MOV, MKV, WEBM and anything else ffmpeg reads — up to 500 MB, 2 hours of audio or 30 minutes of video
Upload a recording and get it back with the dead air taken out — the run-up before you started, the gaps while you thought, the pause where somebody else was talking off-mic. The work is done with ffmpeg on our server rather than in your browser, so the file can be as large as your connection can carry, and a video comes back cut rather than merely muted.
Every tool in this category has to answer one question: below what level is a sound silence? Almost all of them answer it with a constant, usually −50 dB, and some of them make you type it in yourself. That number is only meaningful relative to the noise floor of the recording it is applied to, and we measured what happens when it is not. The same speech, with the same four pauses in it, at four different noise floors:
| The recording's noise floor | A fixed −50 dB threshold | A threshold measured from the file |
|---|---|---|
| Studio-quiet, −90 dBFS | found every gap | found every gap |
| A quiet room, −61 dBFS | found every gap | found every gap |
| A normal room, −46 dBFS | found nothing | found every gap |
| A café, −34 dBFS | found nothing | found every gap |
The bottom two rows are not an edge case — they are a phone on a table and a laptop microphone, which is what most recordings are. And at a −34 dBFS floor the problem cannot be fixed by typing a better number: we tried −50, −40, −35 and −30, and none of them found a single pause, because the entire recording including its silence sits above all of them. We measure the floor instead — the tenth percentile of fifty-millisecond window energies, which is where the quiet parts actually sit — and put the threshold five decibels above it. On those same four files that removed 5.17, 5.29, 5.44 and 5.82 seconds where 5.00 was the known-correct answer.
The margin above the floor is the one remaining free parameter, so it got the same treatment. At +3 and +4 the detector found nothing on any file with a real noise floor — too close to the floor, and the floor's own variation swamps it. At +6 and +8 it started eating quiet speech, removing 7.31 seconds of a file whose correct answer was 5.00. At +5 it landed within 0.15 seconds of the truth on three of the four. There is also a detail behind that which cost a benchmark round: ffmpeg's own silencedetect could not be made to work across codecs at all — on identical content it found four gaps in the WAV and none whatsoever in the MP3 or the Opus, because a lossy codec's silence is coder noise with louder frames in it. We threshold the smoothed RMS series instead, which gives the same gaps to within 0.05 seconds in WAV, AAC, MP3 and Opus alike.
That is the fear with an automatic silence cutter, and it is measurable rather than a matter of taste. We took a clean studio recording, found the cut points using a noisy copy of it, applied those cuts to the clean original, and compared the total speech energy before and after: 100 to 101 per cent survived, at every padding setting including none at all. The words are not what the padding protects. What it protects is rhythm — the short breath inside a sentence that makes speech sound like speech. Gentle removed none of those at all on all four test files, normal removed half a second to seven tenths, and tight removed one and a half. That is the real difference between the three settings, and it is why tight sounds like a podcast edit and gentle sounds like you.
A silence remover that only handles audio is solving half the problem. Cutting a video means removing the same instants from the picture and from the sound, which rules out copying the streams and means the picture is re-encoded. There is a faster way to do it — ffmpeg's concat demuxer, which cuts at packet boundaries — and we tested it properly before rejecting it: a clip that flashes white and beeps at exactly the same instants, cut both ways, then measured flash by flash. The demuxer put picture and sound up to 983 milliseconds apart and turned 45 beeps into 90. The trim-and-concat graph we actually use held them to 17 milliseconds and kept all 45. It is slower. It is the one that works.
Audio goes through a streaming filter, so it is linear and quick: five minutes takes about 5.8 seconds, ten minutes 11.8, twenty minutes 21.6. Video is a full re-encode, so it depends on the picture — about ten times faster than real time at 360p, three times at 720p, and roughly real time at 1080p. That is also why the length limit for video is lower than for audio: thirty minutes of 1080p is already half an hour of work for the server, and two hours of audio is four minutes.
Free use is five minutes a day, from files up to 100 MB. A credit pack raises that to 500 MB per file, two hours of audio or thirty minutes of video — and credits are charged on the length of the file that comes back, not the one that went in, because charging you for the silence we deleted would be charging you for our own output. You get back the format you sent where we can manage it, at a quality where the encoder is not the thing you can hear. Uploads are deleted as soon as the job finishes and the result six hours later.
Anything quieter than your recording's own noise floor plus five decibels, for at least as long as the setting you chose — a second on gentle, half a second on normal, a third of a second on tight. That means genuine quiet in context: the hiss of your room still counts as silence if there is nothing on top of it, while a whisper does not. What it will not do is remove a sound because it is uninteresting. A long chunk of background chatter, a dog, traffic through a window — those are above the floor, so they stay.
Because it is the wrong question to ask you, and the number you would type is unknowable without measuring the file first. We tested this directly: on a recording whose quiet parts sit at −34 dBFS, which is an ordinary phone recording, not one of −50, −40, −35 or −30 dB found a single pause. The whole recording, silence included, is above all of them. Measuring your file's floor and working from that found every pause on every test file, including the studio-quiet ones where a fixed number happens to work. The three presets exist for the decision that is genuinely yours — how aggressive an edit you want — and the decibel value is not that decision.
No, and we measured it rather than asserting it. Taking the cut points from a noisy copy of a recording, applying them to the clean original, and comparing total speech energy before and after gave 100 to 101 per cent surviving — at every padding setting, including with the padding turned off entirely. The padding is not there to save your consonants. It is there to keep the recording from sounding machine-gunned: it leaves a little of each pause behind so the speech still breathes.
Normal for almost everything: it takes out pauses from half a second up and leaves 0.15 s of each one behind. Gentle only touches pauses of a full second or more and leaves a quarter-second — on our test files it removed none of the short pauses inside sentences at all, so it is the setting that cannot change how you sound. Tight starts at a third of a second and leaves 0.08 s, which on the same files removed about a second and a half of within-sentence pause; it makes a recording noticeably faster and noticeably more edited, which is sometimes exactly what you want and sometimes not.
Yes, and the picture is cut along with the sound rather than merely being silenced. That requires re-encoding the video, because you cannot remove an instant from a stream by copying it, so a video takes longer than audio — roughly a third of its running time at 720p and about its own length at 1080p. In exchange the two stay in sync: on a test clip that flashes and beeps on the same schedule, the picture and the sound came out 17 milliseconds apart at the last cut of 45. The faster method we tested instead of this one drifted by almost a second, which is why we are not using it.
Audio: MP3, WAV, FLAC, M4A, AAC, OGG, Opus, WMA and anything else ffmpeg reads. Video: MP4, MOV, MKV, WEBM, AVI and the rest. You get back the same format where we can — an MP3 returns an MP3 at 320 kbps, a WAV returns a WAV, a FLAC returns a FLAC — and MP3 or MKV for the odd formats nothing else handles cleanly. Cutting rewrites the file either way, so there is no lossless path here the way there is with our trimmer.
For audio, seconds: ten minutes of it takes about twelve, and twenty minutes about twenty-two, because the audio path is a streaming filter and scales linearly. For video it depends on the picture — about ten times faster than real time at 360p, three times at 720p, and roughly real time at 1080p — because every frame that survives the cut has to be encoded again. Uploading a large video is often the slower half.
Your upload is deleted as soon as the job finishes, and the shortened file six hours later. That window exists only so the download link still works if you close the tab and come back. Nothing is kept after that, nothing is listened to, and nothing is passed to anyone — worth stating plainly for a tool people bring interviews, lectures and unreleased recordings to.