How do I remove filler words from a video automatically?

The video is transcribed, filler sounds are matched against a fixed lexicon, and the matching word spans are cut from the audio and video together. Repeated takes are found by comparing the words themselves. Caption & Cut does this in the browser and shows every planned cut before rendering.

Does removing filler words need AI?

Not for the common cases. Sounds like uhm and word-for-word repeated sentences are found by code, exactly and reproducibly. A language model is only needed for cases that require understanding, such as a passage re-recorded in different words.

Is it safe to let a tool cut words automatically?

Only if you can see the cuts first. A tool that removes a meaningful word ruins the video, so every planned cut should be visible and reversible before rendering.

Remove filler words from video

"Uhm". The half-started sentence you abandoned. The take you recorded three times because the phone rang. All of it is in the file, all of it has to go, and finding it means listening to your own recording end to end.

Three different problems, often confused

Tools that advertise "filler word removal" usually mean the first of these three. They are not equally hard.

Filler sounds. "Uhm", "äh", "er". A closed set, so a lexicon catches them. One catch that trips up multilingual tools: "um" is a filler in English but an ordinary word in German ("es geht um") and the indefinite article in Portuguese. A lexicon that ignores the language of the recording will quietly delete real words.

Stutters and repeated takes. "The the", or the same sentence spoken twice because you restarted. Found by comparing the words themselves, no model needed. The interesting decision is which occurrence to keep: the second one, because the first is the failed attempt.

Re-recordings in different words. You said it, disliked it, and said it again differently. Nothing matches literally here, so this genuinely needs a language model that understands both versions are the same point.

Why the boring cases should not use AI

It is tempting to hand the whole job to a language model. We tried, and it was the wrong call. Asked for exact word positions, the model was sloppy: off-by-one boundaries, missed duplicates, occasionally a span covering a sentence that should have stayed.

Fillers and literal repeats are now found by code. They are exact, reproducible, and free. The model only sees what actually needs judgement, and its suggestions pass checks first: a claimed re-recording must be followed by a similar passage, or it is discarded.

See the cuts before they happen

This matters more than the detection quality. A tool that removes a meaningful word has ruined the video, and you will only notice on playback. So every planned cut is shown as struck-through text in the transcript before rendering, and one click puts any word back.

In our own tests the model occasionally wanted to cut a sentence that explained something to the viewer, simply because it mentioned the recording. That is exactly the kind of decision you want to see rather than discover.

Doing it with Caption & Cut

Upload the raw file and tick the disfluency option. You get the transcript with every proposed cut marked, and you decide. The same pass also removes the pauses, and burns in word-by-word subtitles if you want them.

What it will not do

It never rewrites your words, it only marks them for removal. It works on talking-head material, not music or multi-camera edits. And the model step is not perfectly repeatable: the same transcript can get a slightly different verdict on a second run, which is why the review step exists.

Try it

First three videos free, nothing to install, runs in the browser. Free results carry a small watermark. Upload a video.