How to remove silence from video without rushing the words

Cut the wait between thoughts. Keep the breath, the meaning and the picture in sync.

Free toolRemove Silence from VideoFind speech, review every proposed pause and choose a natural pace. Keep deliberate pauses, then save picture and sound together.Remove video pauses

Find pauses, review the joins, then save both tracks

Detect the gaps between spoken phrases and remove selected gaps from picture and sound together. Preview the result before exporting. An automatic first pass cannot decide whether a silent moment helps the story.

Open a short recording and start with natural pacing.
Listen around each proposed join. Restore demonstrations, reactions and deliberate pauses.
Increase protection around speech if a quiet opening or ending is clipped.
Save a new video, play it with sound and generate captions for the finished timing.

Speech detection and a volume threshold answer different questions

Speech detection estimates whether a voice is present. A volume threshold asks whether the sound is quiet. Constant fan noise can sit above a volume threshold even when nobody speaks.

Method	Useful when	Limit
Speech detection	A spoken recording has background sound or changing voice levels.	Soft speech, music and overlapping speakers can confuse it. It does not understand words.
Volume threshold	There is a clear loudness gap between the room and the voice.	A high threshold can cut quiet words. A low threshold can retain noisy pauses.
Transcript editing	You also need to remove a repeated sentence or a filler word.	Transcription needs more work and can mishear words. Review the actual sound.
Manual timeline cuts	The silent action or performance needs careful judgement.	You must inspect and cut each gap yourself.

Floi uses Silero VAD locally, with one probability per 32 ms analysis frame. The decision about what to remove is separate from that probability. Do not mistake the waveform’s loudness for the model’s speech estimate.

Keep enough room around the words

Use a minimum gap to avoid editing every breath. Then keep protection around each phrase and, if needed, some extra pause time.

Minimum pause

Only gaps longer than this qualify. Raise it to leave ordinary speech rhythm alone.

Speech protection

Keep a short margin on both sides of the detected voice. It protects soft beginnings and endings.

Extra pause time

Leave some quiet between thoughts. This adds to the two speech-protection margins.

For a detected 2.00-second gap, 0.16 seconds of protection on each side plus 0.30 seconds of extra time leaves 0.62 seconds. The proposed cut is 1.38 seconds. This is a worked timing example, not a speech-accuracy measurement.

Why the picture must follow the sound

Every removed interval must leave both tracks at the same time. Removing sound alone makes the following words arrive before their pictures.

Suppose you keep source ranges 0 to 2 seconds and 4 to 7 seconds. The second range starts at 2 seconds in the edited file. A spoken event at source time 5.5 seconds must appear at edited time 3.5 seconds in both tracks.

Retained source range	Edited range	Time removed before it
0.00 to 2.00 s	0.00 to 2.00 s	0.00 s
4.00 to 7.00 s	2.00 to 5.00 s	2.00 s

Floi rounds cut boundaries to its 48 kHz audio timeline. It draws the picture on a 30 fps output timeline using the same retained intervals. Four-millisecond fades soften sound joins without overlapping or shortening those intervals.

What the synthetic export checks showed

On 31 October 2026, we tested original synthetic speech with two long pauses on an Apple M1 Max with 32 GB memory. Chrome 156 and automated WebKit 27.2 saved the edited files. FFprobe counted frames, while FFmpeg checked the picture and sound against the source.

Check	Chrome	WebKit
Natural cut plan	9.166667 s of source picture became 6.118667 s.	The same two cuts and retained duration.
Saved MP4	184 picture frames at 30 fps. Container duration: 6.133333 s.	184 picture frames at 30 fps. Container duration: 6.133333 s.
Sound after the cuts	No detected offset in three sampled AAC windows.	No detected offset in three sampled AAC windows.
Saved WebM sound	No detected offset in three sampled Opus windows.	Sampled offsets ranged from −5.8 to +4.4 ms.
Rotated and variable-rate inputs	Saved at the intended orientation and 30 fps output timing.	Saved at the intended orientation and 30 fps output timing.

The video’s final frame rounds to a 30 fps boundary, so the saved container can outlast the exact cut plan. Audio correlation used 32 ms windows and a search step of four audio samples. A reported zero means no offset detected at that resolution. It is not proof of perfect timing for every recording.

Four sampled pictures per inspected export matched the expected source times, after conversion to the same pixel format. These checks test timeline correspondence. They do not measure speech-detection accuracy, all-frame perceptual quality or physical phone performance. MP4 and WebM still re-encode the recording.

Choose the right workflow for the recording

A browser studio suits one short spoken clip. A desktop or transcript editor suits long sessions, many tracks or decisions about the actual words.

Option	Useful advantage	Tradeoff
Floi	Local speech detection, pace presets, individually restored pauses and no account to save.	Five-minute/300 MB inputs, a 1920-pixel output long edge and 30 fps re-encoding. No transcript or bad-take selection.
Rescreen	Our short synthetic test reached a result with exact pause review and threshold controls.	Its loudness method needs a useful separation between the voice and the room. Its larger stated limits were not tested.
CoProd	Our clip reached a two-cut review with simple pacing presets and editor timeline options.	No individual restoration control was visible in our test. A finished saved file was not verified.
Gling	A broader transcript workflow includes filler words and bad takes.	Its pricing page lists one hour of free AI-edited media monthly and watermarked free exports. We did not test its editor.
A desktop timeline or FFmpeg	Long recordings, repeatable processing and detailed track control.	Installation and a more involved review or command-line workflow.

Competitor observations date to 31 October 2026. Their features and prices can change. Local processing is available elsewhere too. The useful choice is the workflow that lets you review your recording properly.

Use FFmpeg to inspect silence, then join matched intervals

Use silencedetect to report quiet ranges without editing the source. Adjust its threshold for your recording and check the reported boundaries.

ffmpeg -i input.mp4 -af "silencedetect=noise=-35dB:d=0.8" -f null -

The command below keeps 0 to 2 seconds and 4 to 7 seconds from both tracks. Replace those ranges with intervals you have reviewed. The example requires an input with video and audio.

ffmpeg -i input.mp4 -filter_complex \
  "[0:v]split=2[v0][v1];[0:a]asplit=2[a0][a1]; \
   [v0]trim=start=0:end=2,setpts=PTS-STARTPTS[v0t]; \
   [a0]atrim=start=0:end=2,asetpts=PTS-STARTPTS[a0t]; \
   [v1]trim=start=4:end=7,setpts=PTS-STARTPTS[v1t]; \
   [a1]atrim=start=4:end=7,asetpts=PTS-STARTPTS[a1t]; \
   [v0t][a0t][v1t][a1t]concat=n=2:v=1:a=1[v][a]" \
  -map "[v]" -map "[a]" -c:v libx264 -crf 20 \
  -c:a aac -b:a 192k -movflags +faststart output.mp4

FFmpeg’s silencedetect documentation explains its threshold and duration. Its concat filter requires each segment to start at timestamp zero. The setpts and asetpts steps do that here.

Caption the final timing

Make captions after the pause edits are settled. Old subtitle times do not move automatically when a gap disappears. Burned-in captions move with the picture, but the reading time may become too short.

Play the saved file with sound. Check joins near the beginning, middle and end.
Check quiet words and any retained demonstration.
Use the subtitle generator on the edited file, then correct the words.
Keep the original recording until the final export has been reviewed.