✂️ Silence & Filler Word Remover

What "Silence Detection" Actually Measures

The tool does not listen for speech. It measures loudness, sample by sample, and decides which stretches of the recording are quiet enough to be considered a pause. The audio is decoded into raw samples through the Web Audio API, then scanned in short windows — a few milliseconds each. For every window the tool computes an RMS (root mean square) amplitude, which is a fair approximation of perceived loudness, and compares it against a threshold derived from the Silence sensitivity slider.

This distinction matters because it explains almost every surprising result. A pause where the air conditioner is running is not silent in RMS terms. A whispered word is quieter than a noisy room. If the tool keeps a gap you expected it to cut, the gap is not actually quiet; if it cuts a word, that word was quieter than the background noise floor.

The Three Sliders, and What Each One Trades Away

The default of 400 ms for minimum silence length is not arbitrary. Natural speech contains pauses between clauses that typically run 150–300 ms, and breath pauses that run 300–600 ms. Setting the minimum below roughly 250 ms starts removing the grammatical pauses that make a sentence parseable, which is why aggressively trimmed audio can feel exhausting to listen to even when nothing is obviously wrong with it.

A Worked Example

Take a 12-minute screen recording of a software walkthrough. A typical unscripted take of that length contains somewhere between 90 and 140 pauses longer than 400 ms — reading the next step, waiting for a page to load, thinking about phrasing. If the average removable pause is 700 ms and 110 of them are cut with 120 ms retained each, the arithmetic is:

110 × (700 − 120) = 63,800 ms ≈ 1 minute 4 seconds removed

That is roughly a 9% reduction, and it usually feels like considerably more than 9% because the pauses being removed are exactly the moments where attention drifts. Longer dead air — a 20-second wait while a build compiles — is where the tool earns its keep, because a single such gap is worth twenty ordinary pauses.

Filler Words: What Is and Is Not Possible in a Browser

Removing "um", "uh", "like" and "you know" is a different problem from removing silence. Silence is a signal-level property you can measure directly. A filler word is a linguistic category — you have to recognise the word before you can decide to delete it, which requires speech recognition, and you have to know where its boundaries fall to the millisecond, which requires forced alignment.

In practice, most filler words are surrounded by pauses, so silence trimming already tightens them considerably. Where a filler is welded onto the following word ("um-so-anyway"), no threshold will separate them cleanly, and the honest answer is that this is a manual edit. The review list exists precisely for these cases: you can disable a cut the tool proposed, or leave a region alone and handle it in a full editor afterwards.

Why Manual Review Is Not Optional

Every automatic trimmer produces a small number of wrong cuts, and the failure modes are predictable. Plosive consonants (p, t, k) begin with a brief closure that is genuinely silent — trim into it and the word loses its attack. Trailing fricatives (s, f, sh) decay gradually and can drop below threshold before they have finished. Speakers who drop pitch and volume at the end of a sentence often lose their final syllable.

Toggling individual cuts on and off in the review list is faster than it sounds, because errors cluster: they happen at the same speaker's same habits, so once you have spotted one you know what to look for. Scanning the shortest proposed cuts first tends to surface problems quickly, since a wrongly detected "silence" inside a word is almost always a short one.

When You Should Not Remove Pauses

• Teaching and tutorials. A pause after a new concept is processing time. Removing it does not make the video better, it makes it harder to follow.
• Anything with music or ambience. Beds, room tone and background music make the whole track loud enough that silence detection has nothing to grab; and if it does cut, the ambience will jump audibly at every edit.
• Interviews and multi-speaker recordings. The gap between one person finishing and another starting carries meaning. Compressing it makes people sound like they are interrupting each other.
• Anything that will be dubbed or subtitled later. Timing changes invalidate existing caption files.

Known Limitations, Stated Plainly

Everything happens in your browser, which sets real ceilings. Decoding audio expands it dramatically in memory — an hour of stereo audio at 48 kHz occupies roughly 690 MB as 32-bit float samples — so very long files can exhaust a tab before processing starts. Files in the 5–20 minute range are comfortable on ordinary hardware; feature-length recordings are not the right job for a browser tool.

For video, the trimmed export is produced through the browser's own recording pipeline, so the container and codec are whatever your browser supports rather than a format you choose. The result is fine for review copies, uploads and internal sharing. If you need a specific codec, bitrate or a frame-accurate master, run the trimmed file through a desktop encoder afterwards. The tool also processes a single file at a time — there is no batch queue — and it edits only in time, never adjusting levels, noise or EQ.

Nothing Is Uploaded

The file you choose is read with the browser's File API and decoded locally. No copy is sent to a server, there is no account, and nothing is retained after you close the tab. You can verify this rather than take it on trust: open your browser's developer tools, switch to the Network tab, and process a file — you will see no outbound request carrying your media. A stricter test is to load the page, disconnect from the internet, and use the tool anyway. It keeps working, because there is nothing on the other end to talk to.

This matters more than it might seem for this particular tool. Raw recordings are the least filtered thing most people produce: unedited takes contain false starts, names, internal details and things said in the belief they would be cut. Keeping that material on your own machine is a meaningfully different privacy posture from uploading it to a processing service.

Frequently Asked Questions