How to Remove Filler Words and Bad Takes From a Video

Edit from the transcript, not the waveform. A tool that transcribes every word, including the "um"s, can list each filler word, repeated word and retake with its exact time. You listen to each one, keep what was on purpose, and cut the rest in one go.

In StoryCut: upload your video, open the Remove Silence card, press Find next to Clean up speech, review the list, and press Cut. It takes a few minutes for a video of several minutes, and your captions follow the cuts. Clean up speech is on Pro and up.

Disclosure: StoryCut publishes this guide and makes the tool it describes.

• Filler words: sounds and words that carry no meaning: um, uh, mhmm. Phrases such as "you know", "like" or "basically" are fillers only some of the time: "it was like nothing worked" uses "like" as a filler, "I like this camera" doesn’t.
• Repeated words: a word said twice by accident, as in "and that’s the the part". Some doubles are correct English ("that that", "had had") and should stay.
• Retakes: a sentence you said again because the first try didn’t come out right. The usual rule is to keep the last take, which is the one you meant.
• False starts: a sentence you abandoned, often with a cue such as "sorry, let me start again".

Pauses are a fifth kind, and the oldest one: long silences between sentences. Remove Silence cuts them at the pace your edit style sets.

Each item shows what was said around the cut, how long it is, and why it was found. For a retake, you see the line that goes and the line that stays.

Because the mistakes an automatic cut makes are the ones viewers notice: a repetition you made on purpose for emphasis, a list ("faster, faster, faster"), the wrong take kept, a word clipped at its edge. That is why most tools that do this show the cuts before making them.

StoryCut is careful in a few specific ways:

• A retake is only proposed when what you cut is said again in the take you keep. If the AI isn’t sure, the item is listed but switched off.
• Repetitions the AI judges intentional, like lists and emphasis, aren’t listed at all.
• Retakes can never take away more than a fifth of your speech.
• Phrases like "you know" are only proposed where they mean nothing, and the unsure ones say "Not sure: listen first".

Because transcripts aren’t precise enough for audio. In continuous speech, the timings a transcription gives for each word touch each other and can be up to 90 milliseconds late, so a cut placed on those timings can take the start of the next word with it.

StoryCut places every cut edge from the sound itself: it looks for the quietest moment of the pause nearest the removed word, and adds a 12-millisecond fade at each join so there is no click. When a filler sits in continuous speech with no quiet moment around it, the cut would be harsh, so it is never switched on by default.

Cutting by hand? Zoom into the waveform and cut in the dip between words, never on the word itself, and add a tiny audio crossfade at every join.

Without a tool that hears the fillers, the manual method works in any editor (CapCut, Premiere Pro, DaVinci Resolve, Final Cut):

• Find them: scrub through at 1.5× speed and drop a marker at each "um", repeat and restart. Fillers show up as small isolated blobs on the waveform between longer words.
• Cut retakes first: they are the biggest wins. Keep the last take, and listen to the sentence before it to make sure the join makes sense.
• Then fillers: cut in the silence on both sides, and add a short crossfade.
• Cover jump cuts: on a talking head, a cut makes the face jump. A small zoom on the next shot hides it, which is what Magic Zoom does automatically.

By hand, that means listening to the whole video at least once. With a review list, you only listen to the cuts.