Cut the mistakes out of a talking video: retakes, filler words and pauses
Retakes, false starts, fillers and dead air are four different cuts. How to ask for the ones you want, which take wins, and how to read the result as a script.
You pressed record, talked for four minutes, started one sentence three times, lost your place twice and said “um” more than you would like to count. Somewhere in there is a good ninety seconds. Finding it by hand means listening to yourself on a loop and trimming at the level of a syllable, which is why most takes like this never get posted. It is also the kind of edit an AI editor does best, because everything it needs is in the words. This is how to ask Narrative to cut the mistakes out of a talking video, what it does at each cut, and how to check the result.
Why this works: the editor reads the take
When your clip arrives it is transcribed, and every word gets a start and an end time, to the hundredth of a second. From then on the editor is not looking at a waveform and guessing where the pauses are. It is reading a script that happens to have timestamps, and the mistakes are visible in it the way they would be on a page: the sentence that begins twice, the word said two times in a row, the “so, um, yeah”, the four seconds of nothing while you looked at your notes.
That is also why you do not need to mark anything. You do not scrub for the bad parts and list their times. You say what counts as a mistake and what you want left.
What to ask for
A request that works on almost any take:
Cut out the mistakes: retakes, false starts, filler words and long pauses. When I say a line more than once, keep the last full version. Keep everything else in the order I said it. Add captions.
Four things in there are doing work:
- Name the kinds of mistake. “Clean it up” leaves the editor to decide how clean. Retakes, false starts, fillers and pauses are four different cuts, and you may not want all four.
- Say which take wins. When a line is said three times, the last complete one is usually the one you were happy with, since it is the one you moved on from. If your habit is the opposite, say “keep the first”. If one take is better for a reason the words do not show, name it: “keep the version where I laugh”.
- Say what must stay. “In the order I said it” stops a tidy-up turning into a restructure. If you would welcome a restructure, ask for one separately.
- Give a length only if you have one. “Under 60 seconds” is a different job from “remove the mistakes”: it means good material goes too. Both are fine requests. Decide which one you are making.
What happens at each cut
A cut in the middle of someone talking is easy to make and hard to make invisible. These are the rules the editor works to.
- Cuts sit on word boundaries. Not “around 0:14”, but at the onset of the first word kept and the end of the last. A filler in the middle of a sentence can come out without clipping the words on either side.
- The dead air goes. Kept stretches are butted together. A cut that leaves half a second of silence at every join looks like footage nobody edited.
- Sentences stay whole. A clip does not end mid-sentence or begin mid-clause. A dangling “and…” either runs on to the end of its sentence or the cut moves back to before it.
- The sound overlaps slightly. When the same voice continues across a join, the next line begins about a tenth of a second before the previous word has finished ringing. Without that, every join has a tiny hole in the audio and the ear hears the edit even when the eye does not.
- The cuts stay hard. No dissolve between two halves of a sentence. A straight cut on a talking head reads as pace; a crossfade reads as a mistake.
The jump in the picture
Take words out of a single continuous shot and the picture jumps at each cut: your head is in a slightly different place. On short-form video this is normal, and viewers read it as energy. If you would rather soften it, there are three honest options, and each is one sentence:
- Punch in. “Alternate between the normal framing and a slightly closer one at each cut.” The change of size makes the jump look deliberate.
- Cover it. “Put the screen recording over the cut at 0:20.” A second shot over the join hides it completely while your voice carries on underneath. This needs a second shot, so add one to the project: a clip of the thing you are talking about, a photo, a screen recording.
- Loosen it. “Leave a short breath between sentences.” Fewer, softer cuts, at the price of a slightly longer video.
How tight is too tight
With every pause removed, a person sounds like a list being read out. The pause before a punchline and the breath after a hard sentence are part of what you said. If the first pass feels breathless, say where it should relax, not just that it should:
- “Keep the pause before ‘and that is when it broke’.”
- “Tight in the first ten seconds, more relaxed after that.”
- “Leave the fillers in where I am thinking out loud at the end; it sounds more like me.”
Check it in two minutes
- Watch it once with the sound on, at normal speed. You are listening for a clipped word or a join that sounds swallowed. Note the time of anything that catches.
- Read the captions as a script. On the iPhone, tap a caption on the timeline and then Edit. Every phrase of the cut is listed under its time. Reading it top to bottom takes seconds and shows at once whether a retake survived or a sentence lost its ending. Tap any line to jump the video there.
- Fix the words while you are there. Type over a misheard name in that same list; the correction keeps its timing. Delete removes a caption that should not be on screen at all.
- Send back what you found, with times. “At 0:31 the word ‘really’ is cut off; give it a little more room. At 0:48 I say ‘basically’ twice, cut the first.” Two specifics get fixed in one pass.
For a single join that is a few frames out, doing it yourself is faster than describing it: select the clip, drag its handle, done. The guide to fixing an edit by hand covers trim, split and delete on the timeline.
The same goes for sound. If the music sits too loud under one sentence, or a clip’s own audio needs to come down, select the clip or the audio track on the phone and tap Volume: one slider, from silent to twice the original level, with a mute button beside it. The slider shows the level the editor set, so you are adjusting its mix, not starting over.
Captions come with it
Plenty of people watch a talking video with the sound off, so ask for captions in the same request. They are drawn from the same word timings as the cuts, which means they only contain the words that survived: a retake you removed cannot reappear as text. When the cut changes later, the captions re-form around it without being asked.
The look is chosen for you from the footage and the request, and it is a tap to change. The caption styles guide shows the seven looks on the phone and how to switch between them.
What it cannot fix
- A line you never said cleanly. If all three takes of a sentence stumble, the editor can pick the least bad one or cut the sentence. It cannot assemble a clean take from words you did not say. Record that line again and add the clip.
- Sound that was bad at the source. Wind, a distant microphone and a loud room are still there after the cut. Tight editing makes a noisy take shorter, not cleaner.
- A change of mind. If you said something in minute one and contradicted it in minute three, both are fluent and neither is a mistake. Say which you meant.
Longer recordings
The same request works on a twenty-minute recording, but there the better question is usually which ninety seconds to keep, not which pauses to cut. That is a different job with its own guide: turning a long recording into short clips.
Try it on the take you gave up on
Everyone has one in their camera roll: the take that was nearly right and not worth the hour to fix. Open the iPhone app, add that clip, and paste in the request from the top of this page. Then read the captions as a script, and send back the one or two things you would change.
