Filler word remover

What this tool does

Spoken English is full of small words that hold the floor while the speaker thinks: um, uh, you know, I mean, sort of. Nobody notices them in conversation. They are impossible to ignore on the page. A ten-minute interview transcribed verbatim can carry two hundred of them, and each one drags the reader out of the sentence.

This page removes them. You paste the transcript, tick the tics you want gone, and the cleaned version appears underneath with a count of exactly what was taken out. The text never leaves your browser — there is no upload step and no server call, so a confidential interview or an unreleased product discussion stays on your machine.

Why some fillers are switched off by default

A filler word remover that blindly deletes every instance of like will happily turn “a tool like this one” into “a tool this one”. That is worse than the original problem, because the damage is silent: nothing on screen tells you a real word was destroyed.

Three of the toggles start switched off for exactly that reason.

• like — a preposition and a verb before it is ever a tic. “I like it”, “something like that”, “it works like a shortcut”.
• actually — often load-bearing. “It actually shipped” means something different from “It shipped”.
• right — collides with “turn right”, “that’s right” and “the right column”. As a tic it is almost always the tag question “…, right?”, which is easier to catch by eye than by rule.

Turn them on when you know your transcript, read the result, and turn them off again if the count looks too high. The per-word tally under the output exists so you can make that judgement in a couple of seconds rather than re-reading the whole file.

How the matching works

Every filler is matched on word boundaries and without regard to case, so likely, actually inside factually, ah inside ahead and um inside umbrella are never touched. Multi-word phrases such as you know and kind of tolerate any amount of whitespace between the words, including a line break, and are removed before the single-word rules run so neither half can be eaten first.

Removal always leaves a mess behind, and cleaning it up is most of the work. A filler is usually parenthetical, so the match reaches out to the commas on either side of it: “we, um, should ship” becomes “we should ship” rather than “we, should ship”. Whatever is left is then repaired — doubled commas collapse, a comma that has drifted in front of a full stop is deleted, the space that used to sit before punctuation goes, runs of spaces squeeze down to one, and a full stop orphaned at the start of a line by a removed “So yeah.” is dropped. If you leave “Re-capitalise sentence starts” ticked, the capital that vanished with a leading “Um,” is put back on the following word.

What it deliberately will not do

• It does not rewrite your sentences. Removing filler is a subtraction, not a rephrasing — the words that remain are the words you said.
• It does not delete repeated words (“the, the main thing”). Stutters and genuine repetition look identical to a regular expression, and guessing wrong changes meaning.
• It does not remove [inaudible], [crosstalk] or timestamps. Those belong to the transcript cleaner, which handles transcript furniture as separate passes.

A sensible order for a raw meeting transcript

If you are working from an export out of a meeting recorder or captioning tool, run the structural cleanup first and the vocabulary cleanup second. Timestamps and speaker labels confuse sentence-level rules, and repairing punctuation twice is wasted effort.

• Strip timestamps, speaker labels and hard wrapping in the transcript cleaner.
• Paste that result here and remove the filler words.
• Check the length against your target with the word counter, or against a time limit with the speech time calculator if the text is going to be read aloud again.

For subtitles, convert the file to plain text with the subtitle converter first — cue numbers and timing lines will otherwise be treated as part of the prose.

Common questions

Is my transcript uploaded anywhere?

No. The matching and the cleanup both run in JavaScript inside this tab. There is no request to a server when you paste, type or copy, which you can confirm in your browser’s network panel. Closing the tab discards everything.

How many words should I expect to lose?

Unscripted speech usually runs between three and eight per cent filler; a nervous or unrehearsed speaker can go past ten. If your count comes back near zero the transcript has probably been cleaned already, and if it comes back above fifteen per cent check whether you switched on like or right and caught real words with them.

Can I add my own filler word?

Not in this version — the list is fixed so the matching rules can be tuned for each entry rather than accepting arbitrary input. If a word is doing real damage to your transcripts, tell us which one and we will look at adding it with the right boundary rules.

It ate a comma out of my list.

That is the one known cost of treating fillers as parenthetical. “Eggs, milk, um, bread” comes back as “eggs, milk bread”, because the commas that wrapped the um look identical to the commas that separate list items. It is rare enough — and visible enough in the side-by-side output — to be worth the much cleaner result everywhere else. Put the comma back by hand when it happens.

Why does the word count drop by more than the filler count?

An em dash or a stranded comma left hanging by a removal is tidied away with the filler, and a dash counts as a token when words are split on whitespace. The tally under the output is the accurate number: it counts matched filler words, not incidental punctuation.

Better still: never dictate them in the first place

Cleaning a transcript is a repair job. VoiceSnap Pro is the version where the repair is unnecessary: you hold one keyboard shortcut, speak, and clean punctuated text appears in whatever field your cursor is already in — your email client, Slack, a code editor, a browser form. Filler words are stripped as you speak, punctuation and paragraph breaks are added automatically, and every dictation is kept in a searchable notes library in case you need the original wording back.

It runs on macOS and Windows, handles 50+ languages including mid-sentence switching, takes a custom vocabulary for names and jargon that generic models mangle, and is a one-time purchase rather than a subscription. Nothing you dictate is used to train AI models. The app has not shipped yet — join the waitlist and you will hear on release day.