Filler words are not the enemy. Repetition is.

The first thing everyone does with a transcript editor is hunt for “um”. It is satisfying, it is fast, and after about a minute the result sounds wrong. The speaker is still there, but something human has gone missing.

Here is the rule that took us longest to learn: filler words are punctuation. Repetition is noise.

What filler actually does

An “um” before a number tells the listener the number is about to matter. A “like” in the middle of a sentence marks a shift in register from explaining to confiding. A “so” at the start of a sentence is a paragraph break you can hear.

Strip them all and the speech loses its paragraphs. Listeners cannot say what is wrong, they just report that the person sounds “read” or “AI”. Leave a few in and the same recording sounds like someone thinking in real time, which is the whole appeal of a talking-head video.

What repetition does

Repetition is different. When a speaker starts a sentence twice, the listener hears the false start, discards it, and hears the real one. That discard costs attention every time. The same goes for restated clauses, for “what I mean is”, for the second attempt at a word that came out wrong.

None of that carries meaning. All of it can go.

The rule

Strike a word or a phrase when the sentence means exactly the same thing without it. Keep it when removing it changes the rhythm of the sentence.

In practice:

• Strike: false starts, restated clauses, the second “the the”, a word said twice for no reason, “sort of” and “kind of” when they hedge nothing.
• Keep: a single “um” before a fact, a “so” that opens a new thought, a “like” that introduces an example, a laugh.
• Judgement call: “basically”, “actually”, “literally”. Strike them if the speaker uses them every sentence, keep one if it lands.

A rough count from our own edits: about two thirds of filler words stay, and nearly all repetition goes. The result sounds like the speaker on a good day.

How to do it fast

Read the transcript instead of listening to the video. Repetition is hard to hear and easy to see; the eye catches a doubled phrase in a fraction of a second. Strike the whole false start with one swipe rather than word by word, and let the pause dial take care of the gap the strike leaves behind.

Then play it back once, at full speed, without reading. If you notice the cuts, one of them landed on a pause that was doing work. Put that word back. The tap that struck it is the same tap that restores it.