1. Two kinds of filler
Not every filler word is filler. There are mechanical hesitations — um, uh, er — which carry no meaning and can always go. And there are discourse words like 'like', 'you know' and 'I mean', which sometimes fill space and sometimes do real work in a sentence. Removing the first category is always an improvement; removing the second indiscriminately changes how you sound and occasionally changes what you said.
- Mechanical: um, uh, er, mm. Always removable.
- Discourse: like, you know, I mean, right, so. Sometimes structural, sometimes filler.
- Verbal tics that are part of your voice — cutting all of them can make you sound like someone else.
2. What to cut automatically
The safe automatic pass targets hesitations, stutters and immediately repeated phrases, all of which are identifiable from the words themselves and none of which carry meaning. This runs alongside silence removal in a single pass, which is why tightening pacing and removing filler are usually the same request. Because the cuts are placed on word boundaries, what remains still sounds like continuous speech rather than a splice.
3. Keeping the breath in
The thing that makes aggressive filler removal sound artificial is not the missing words, it is the missing air. Speech includes breath, and a version with every inhalation removed sounds like it was assembled rather than spoken. Leave the breaths at sentence ends and section boundaries; remove the ones that precede an abandoned start. Gentler pacing modes preserve this automatically.
If a cleaned-up recording sounds oddly inhuman, it is usually breath rather than words. Ask for more breathing room rather than putting the filler back.
4. Stutters and restarts
Repeated words at the start of a sentence, and sentences begun three times before they land, are the largest single source of clutter in unscripted speech — larger than um and uh combined. They are also the most satisfying to remove, because the result reads as fluency you did not actually have. Keep the final attempt, discard the earlier ones, and the delivery tightens dramatically.
- Repeated first words: keep one.
- Abandoned clauses: cut entirely, including the pause after them.
- Three attempts at the same sentence: keep the clearest, which is usually the last.
5. How far to take it
There is a point where cleanup stops helping. A completely filler-free, breath-free, gapless delivery sounds processed, and on conversational content it removes the very thing that made it feel genuine. Match the intensity to the format: short-form can take the aggressive version, an interview should keep most of its texture, and anything personal should keep nearly all of it.
- Short-form: remove everything. The format expects gapless delivery.
- Explainers and tutorials: remove hesitations, keep breath and section pauses.
- Interviews and personal stories: remove only the obvious stumbles.
✖ Removing every instance of 'like' and 'you know'
Why it fails: These often carry structure or tone, and stripping them all makes the delivery sound unnatural and unlike you.
✔ Better approach: Let the mechanical hesitations go automatically and leave discourse words alone unless they are clearly padding.
✖ Cutting all the breaths
Why it fails: Speech without air sounds synthetic, however clean the words are.
✔ Better approach: Use a gentler pass, or ask for more breathing room between sentences.
✖ Applying maximum cleanup to an interview
Why it fails: Hesitation is part of how a thoughtful answer reads; removing it makes a considered speaker sound glib.
✔ Better approach: Clean only the stumbles and leave the rhythm of the conversation intact.
Frequently Asked Questions
Common questions around this editing workflow.