Removing Filler Words Without Making Speech Sound Chopped

There are two ways to get filler removal wrong, and the less obvious one does more damage. Leave too many hesitations and the speech sounds unprepared. Remove all of them and it sounds edited — clipped, breathless, subtly inhuman, in a way listeners register immediately and describe as the audio sounding "off" rather than as anything they can point at.

The difference between the two is not how aggressive you are. It is knowing which sounds carry information and which do not, and repairing the gaps left behind properly. This guide covers what to cut and what to leave, why breaths are structural, the room-tone patch that makes an edit inaudible, and how to keep the picture from exposing the work you did on the sound.

Fillers Are Not One Thing

Grouping every um, so, like and you know under one heading is what produces over-editing. They do different jobs, and only some of those jobs are dispensable.

The distinction that saves the most work is between hesitation and marking. Um in the middle of a clause is almost always removable. So at the start of a sentence is frequently doing real structural work, and cutting it produces prose that reads correctly and sounds wrong.

Breaths Are Structure

The most common single error in this work is deleting the breath before a sentence along with the filler that preceded it.

An intake of breath tells the listener something is coming. Remove it and the sentence begins from nothing, which sounds like a splice even when the join itself is technically clean. It also removes the natural pause that lets the previous idea land.

Shorten rather than delete. If a breath is long or noisy, trim it and lower its level a few decibels — it stays present and stops being distracting. Where a breath has to go entirely, take the pause down with it and let the sentences run together, so the result reads as fast speech rather than as a missing moment.

The exception is the deep breath at the end of a passage, where the speaker has finished and is resetting. That one can usually go, because there is nothing after it that it was preparing for.

Cut Where the Waveform Is Quiet

Two mechanical habits make the difference between edits you can hear and edits you cannot.

Cut near zero. Audio is a waveform swinging above and below a centre line. Cut it at a point where it is far from that line and the level jumps instantly at the join, which reproduces as a click. Zoom in far enough to see the waveform shape and place cuts where it crosses the centre.

Crossfade everything. One or two frames is enough. It smooths any residual discontinuity and it costs nothing. Applying a short crossfade across every audio edit in a sequence takes seconds in most editors and eliminates an entire category of problem.

Both are quicker than they sound, and neither requires listening to each cut individually — which matters when a fifteen-minute interview has two hundred of them.

Patch the Gap With the Room

Removing a filler leaves a hole, and the hole is not silent in the way you might expect: it is more silent than the recording around it. Digital zero has no noise floor at all, and against a recording that does, the sudden absence is audible as a small drop-out.

The fix is room tone — the ambient sound of the actual space with nobody talking. Lay it underneath the dialogue track for the whole piece so the noise floor is continuous, and every gap you create is filled by something with the same character as its neighbours.

Record thirty seconds of it before leaving any location. Where that did not happen, harvest a second or two from a pause in the recording itself and loop it: same room, same microphone, same floor. What does not work is a generic ambience from a library, because its noise character will not match and the mismatch is as audible as the gap it was covering.

The Picture Has to Survive It

Every clean audio cut in a single-camera piece produces a visible discontinuity: the speaker's head is in a different place either side of the join.

There are three ways to handle it. Cover the join with a cutaway, which is the most common and the reason cutaways exist in the quantity they do — the selection criteria are in choosing B-roll that earns its place. Change the framing across the cut with a punch-in, so the jump reads as an intentional change of shot. Or leave it visible and let the jump cut be part of the style, which works in short-form and rarely elsewhere.

The framing-change threshold matters here: a punch-in that only scales slightly still looks like the same shot with a glitch in it. How much change is needed before a join reads as intentional is set out in the four cuts and where each one belongs.

Plan this while you cut the audio rather than afterwards. An editor who removes two hundred fillers and then goes looking for two hundred cutaways has created a much larger job than one who decided, per join, whether the cut was worth the cover it would need.

How Far to Go

The honest answer is that it depends on what the video is, and the two ends are clearly wrong.

A scripted explainer can take heavy filler removal, because the delivery is meant to sound prepared. A conversational interview should keep most of its texture, because the appeal is that a person is talking. A personal story keeps nearly all of it, because the hesitations are the content.

One useful check: play back the edited passage and ask whether it still sounds like the person who recorded it. If it sounds like a better speaker than they are, you have gone past the point where the edit serves the material.

If the same speech is also being captioned, decide once whether the captions follow the finished edit or the raw take — they have to agree with what the viewer hears, and reconciling them afterwards is more work than deciding up front, as set out in a systematic way to fix bad auto-captions.

Next Steps

Take a two-minute stretch of your last interview and edit it twice: once removing everything you can, once removing only hesitations and repeated words. Listen to both on headphones. Most people are surprised by how much worse the aggressive version sounds.

Then add a continuous room-tone layer under the whole track before you do anything else. It is a two-minute change that removes the most audible artefact of this entire process. Where the recording also needs noise, level and tonal work, do that in the right sequence — the order to clean up bad audio. Everything around it lives in the editing cluster.