Open the recording in the converter. When the waveform is finished, a Remove silence panel appears under the selection controls. No file of your own? Press Open the sample recording, pick the Microphone track and the panel lists 6 stretches, taking 0:20 down to 0:16.
How to use the panel
- Leave Longer than at 1 s. That keeps the short breath between sentences, and a talk still sounds like a talk.
- Leave Quieter than on Auto. The panel prints the limit it chose, such as -48 dB.
- Read the sentence it shows: how many stretches, how much time, and the length before and after.
- Press Remove. Each stretch becomes a row in the cut list, so you can bring back any single one with its Undo button, or press Ctrl+Z to bring them all back.
- Press Play to hear the joins. Preview skips the cuts, so what you hear is what gets saved.
- Convert, then Save file.
The file does not leave your device for any of this. The panel works from the peak levels already computed to draw the waveform, so finding silence takes no second pass over the file.
The test
We built a 6:01 recording from 18 passages of a written lecture spoken by the Windows text to speech voice, with gaps of known length placed between them: 14 stretches of a second or longer, including a 6 s lead-in, an 8 s tail and pauses of 45 s and 90 s. Spoken time above -40 dB was 109.63 s. We then laid steady pink noise under it at six levels to stand in for a room, encoded each as MP4 with AAC at 128 kbps, ran them through the shipped tool in desktop Chrome with the settings above, and saved MP3. Each saved file was then read back with the FFmpeg silencedetect filter as a second opinion.
| Background peak | Auto limit | Stretches | Time removed | Voice lost | Saved MP3 |
|---|---|---|---|---|---|
| Digital silence | -60 dB | 19 | 215.66 s | 0 s | 3,501,096 bytes, 145.9 s |
| -64.7 dB | -58 dB | 19 | 215.66 s | 0 s | 3,501,096 bytes, 145.9 s |
| -54.9 dB | -48 dB | 19 | 216.00 s | 0 s | 3,493,032 bytes, 145.6 s |
| -44.4 dB | -38 dB | 19 | 216.58 s | 0 s | 3,479,208 bytes, 145.0 s |
| -34.7 dB | -30 dB | 19 | 217.33 s | 0 s | 3,460,776 bytes, 144.2 s |
| -28.9 dB | -30 dB | 68 | 88.50 s | 0 s | 6,552,744 bytes, 273.0 s |
| -22.6 dB | -30 dB | None found | 0 s | 0 s | Nothing to remove |
Voice lost counts the spoken time above -40 dB, measured on the clean speech track before noise was added, that fell inside a cut. It was zero in every row, and FFmpeg found no pause of a second or more left in any saved file. The 19 stretches are the 14 we placed plus five pauses inside the spoken passages. With Longer than at 0.5 s the same recordings gave 32 stretches and between 219.00 s and 221.99 s removed, still with no voice lost in the five quieter rows. The tool leaves 0.25 s of the pause on each side that touches speech, which is why the time removed is a little under the total length of the pauses.
Where it stops working
The two lower rows are the limit. Auto never goes above -30 dB, because above that it starts to risk quiet speech. At -28.9 dB the noise only dips under the limit part of the time, so the panel finds 68 short pieces and removes less than half of the pause time. At -22.6 dB it finds nothing and says so. A loud fan, an air conditioner next to the microphone, or music under the talk will end up in those rows, and this tool has no noise reduction to get you out of them.
The test voice is also cleaner than a person. It does not trail off at the end of a sentence, and the noise was steady, where a real room has doors and coughs. A cough counts as sound, so the pause around it stays. A soft last word below the limit could be clipped, which is what the 0.25 s margin and the listen in step 5 are for.
What changes in the saved file
- Cuts are joined with no crossfade. On a real room recording you may hear the background stop and start at a join.
- The output is audio only. The video is not touched, so a saved file with the silence removed no longer lines up with the slides in the original recording.
- The audio is encoded again. Any cut rules out the unchanged M4A copy, so MP3 and AAC are written at 192 kbps. In this test the whole recording as MP3 would be about 8.68 MB and the shortened one was 3.50 MB. The smaller audio guide has the rule for when a cut saves space and when it only saves time; here 40% of the running time stayed, well under the 67% the rule asks for with a 128 kbps source.
- Waveform resolution depends on length. The panel reads about 12 levels per second up to 83 minutes. On a three hour file it is about 5.5 per second, so edges land up to 180 ms further from the speech.
If your file has no decoder in this browser, no waveform is drawn and the panel does not appear; the troubleshooting guide covers that case. A file with several audio tracks is read one track at a time, so pick the microphone track first, as in the screen recording guide.