ThreeCutBlog

← All posts

How to remove filler words and false starts from a video

· 7 min read

A cream paper audio waveform with a gap in the middle, three coral pieces cut from it floating away and coral paper scissors below

To remove filler words from a video, get a transcript with word timings, mark every um, uh, false start and long pause, then cut from just after the word before to just before the word after. Leave a few hundredths of a second either side so no consonant is clipped, smooth each join with a very short crossfade, and shorten long pauses to a beat instead of removing them completely. Done this way, the speaker sounds like themselves on a good day rather than like a robot.

In the example further down, a take loses half its length in this edit and none of its content.

What counts as filler, and what to do with each

Not everything that sounds like filler should go. Sort it into five groups before you cut anything.

What you hear Examples What to do
Hesitation sounds um, uh, erm, hmm Cut every one
Cue words at the edges "Okay, so..." at the start, "Yeah." or "Cut." at the end Cut them; they were said to the camera, not the viewer
False starts "We built this to... we built this so you can..." Cut the abandoned attempt, keep the finished one
Verbal habits like, you know, basically, I mean, so Cut only if the sentence means the same without it
Long pauses A second or more of silence between sentences Shorten to about a third of a second

The verbal habits are where automatic tools go wrong. "It was, like, a mess" loses nothing without "like". "It looked like a receipt" loses its meaning. "So" at the start of a sentence is usually filler; "so we cut the price" in the middle of one is doing a job. Read each one in context before you cut it.

Step by step

1. Get a transcript with word timings

Almost every editor and many phone apps can now generate a transcript where each word knows its start and end time. Editing from text is far faster than scrubbing a waveform: you read the take, and the timeline follows. Before you trust it, fix product names, people's names and numbers, because transcription gets those wrong most often.

2. Mark first, cut second

Read the whole transcript once and mark every filler, false start and pause longer than about half a second. Don't cut yet. Marking first shows you the pattern, such as stumbling only at the start of a take or only before numbers, which tells you what to fix next time you record.

3. Cut on word boundaries, with a little room

Start each cut just after the last sound of the word before, and end it just before the first sound of the word after. Then give the kept words a little room:

  • About 0.05 to 0.1 seconds before a word. Words that start with a breathy sound (h, f, s) begin earlier than the waveform suggests. Cut too tight and "help" becomes "elp".
  • About 0.2 seconds after the last word of a line. Endings trail off. A clipped ending sounds like the speaker was interrupted.
  • Never inside a word. If the filler runs into the next word ("um-so"), keep both or cut both.

4. Smooth the joins

Two cuts that are perfectly placed can still click, because the waveform jumps from one level to another. A crossfade of a few frames (around a tenth of a second) on the sound fixes it.

The picture needs the same care. Cutting the ums out of a talking head video creates jump cuts, where the speaker's head shifts slightly at every join. You have three options:

  • A short dissolve. Four frames at 30 frames per second (about 0.13 seconds) blends the two positions so the cut reads as smooth rather than as a jump.
  • A punch-in. Zoom in 10 to 15 percent on every other section, so each cut looks like a deliberate change of framing.
  • A cutaway. Cover the join with a picture of what the speaker is talking about. This works best where the line describes something the viewer could see.

5. Shorten pauses, but keep one

Cut each gap between sentences down to a beat, about a third of a second. That is enough to separate two thoughts without leaving the viewer waiting. Then put one longer pause back, right before the line the video exists for. A short silence tells the viewer the next sentence matters.

6. Check it the way it will be watched

Listen once with your eyes closed, at normal speed. Any cut you can hear is too tight or needs a longer crossfade. Then watch it on your phone, full screen, with the sound off: if the captions are still showing an "um" or a repeated phrase, the transcript and the cut disagree. (How to add word-by-word captions covers caption timing.)

A worked example

Here is a raw take of about 18 seconds, 36 words:

[1.5 seconds of silence] Okay, so. Um, the, the reason most, the reason most invoices get paid late is, uh, nobody owns the follow-up. [1.2 second pause] Like, you send it and then, you know, everyone assumes someone else will chase it. Yeah.

And the edit, about 9 seconds and 24 words:

The reason most invoices get paid late is nobody owns the follow-up. [beat] You send it, and then everyone assumes someone else will chase it.

What came out, in order:

  1. The silence before the first word.
  2. "Okay, so." and "Um," which were cue words and a hesitation.
  3. "the, the reason most," a false start said twice before the speaker found the sentence.
  4. "uh" in the middle of the line.
  5. Most of the 1.2 second pause, shortened to a beat.
  6. "Like," at the start of a sentence, where it carried nothing.
  7. "you know," which also carried nothing.
  8. "Yeah." at the end, said to the camera.

Every idea in the raw take is still there. A viewer who would have scrolled during "Okay, so. Um, the, the..." now hears the point in the first two seconds.

Prevent most of it while recording

The cleanest edit is the one with the least to cut. Three habits cut fillers at the source:

  • Shorter lines. People stumble at the start of long sentences. A script written in lines of 8 to 20 words, one thought each, gives you far fewer false starts. How to write a video script shows how to break a script into scenes.
  • Record one line at a time. Read the line, look up into the lens, say it, stop. A fumble costs you one line, not the whole take.
  • Start a beat late. Breathe, wait half a second, then speak. Most "okay, so" comes from starting to talk before you've settled.

How to do this in ThreeCut

In ThreeCut you record your video scene by scene in the browser, with each line on screen. After you record, it transcribes every take and plans the cut from the words: it trims the set-up before your first word and the reach for the stop button after the last, drops cue words such as "okay" and "alright" at the start and end, removes ums, uhs, erms and hmms anywhere in the take, removes a phrase you restart straight away, and shortens pauses longer than about 0.3 seconds to about a third of a second. Each join gets a four-frame dissolve, and the captions follow the edited words.

It leaves "like" and "you know" alone, because they sometimes carry meaning; if one bothers you, retake that scene. It only cuts takes that follow their script closely, so a take that wanders off-script is used whole. See the edit on the home page.

Common questions

How do I remove ums from a video automatically?

Use an editor that transcribes the audio with word timings and can delete words from the transcript, so removing "um" from the text removes it from the video. Review the result before you export: automatic tools sometimes cut words that carried meaning, or leave a click at the join that needs a short crossfade.

Should you remove every filler word?

Remove every hesitation sound (um, uh, erm). Remove habits like "like" and "you know" only when the sentence means the same without them. In a long interview a few fillers sound natural; in a 30 or 60 second video each one is a moment where nothing happens.

How long should a pause between sentences be?

About a third of a second is enough to separate two sentences in short-form video. Keep one longer pause, up to a second, before the most important line. Anything longer than that, without a reason, reads as dead air.

Read next