Filler word remover for transcripts

What counts as a filler word

Filler words are the sounds a speaker makes while thinking. VoiceGem ships twelve of them: uh, um, uhm, umm, uhh, uhhh, hmm, hm, mmm, mm, mh, ehh. The list is short on purpose. Longer lists start deleting words that carry meaning — "like", "you know", "I mean" and "right" are all filler in some sentences and content in others, and a filter that removes them silently rewrites what someone said.

Each word is matched at a word boundary, so "umbrella" and "hummus" survive a list containing "um" and "hmm". Matching is case-insensitive, so "Um" at the start of a sentence goes too. A trailing comma or period attached to the filler word is removed with it, which is the detail that separates a clean result from one full of stranded punctuation like ", ," in the middle of a sentence.

A word left in the list that appears nowhere in the transcript costs nothing, and an empty list is a no-op rather than an error. Both behaviors are ported from the app rather than invented for the browser.

How to remove filler words by hand

Find-and-replace is the obvious approach and it fails in two specific ways worth knowing before you spend twenty minutes on it. Replacing "um" with nothing turns "umbrella" into "brella" unless the search is set to whole words, and most editors default to substring matching. Replacing " um " with a single space misses "Um," at the start of a sentence and "um." at the end.

The regular-expression version handles both, at the cost of writing one expression per word and remembering to escape anything unusual. That is the exact loop the tool above removes: twelve expressions, each anchored at a word boundary, each allowing one trailing comma or period, applied in order, followed by a whitespace collapse.

Cleaning a transcript after the fact is the slow version of the job either way. VoiceGem on macOS 14.4 or later runs this filter as the text is produced, so the filler never reaches the document — the browser tool exists to show you exactly what that filter does before you install anything.

Where filler word removal goes wrong

Deleting filler changes the record, and for some transcripts that is not acceptable. A verbatim transcript for a legal deposition, a research interview coded for hesitation, or a clinical note where speech patterns are the observation must not be filtered at all. The safe rule is that filler removal belongs to dictation, where you are producing text, and not to transcription, where you are recording what happened.

The second failure is a filler word that is also a real word in the transcript's domain. "Mm" is a hesitation and also a unit of length. "Hm" is a hesitation and also a chemical symbol. VoiceGem makes the list editable rather than guessing, and the tool above lets you remove any default and add your own before running it.

The third is subtler and shows up in long documents. Removing filler tightens sentences but leaves the sentence structure a speaker produced while hesitating, which reads as abrupt rather than clean. The paragraph formatter on this site handles the structural half of that problem; filler removal only handles the words.

What the browser tool does not do

The filter above is one stage of a longer pipeline, and the page would be misleading if it implied otherwise. Before filler removal, VoiceGem strips paired tag blocks and bracketed artifacts — the "[BLANK_AUDIO]" and "(upbeat music)" that speech models emit on silence. After it, a punctuation mode and an optional lowercase pass run. The punctuation cleanup tool on this site covers those stages.

And the tool cannot do the part that matters most: it takes text you already have. VoiceGem takes your voice, transcribes it on-device, and applies this filter before the text reaches your cursor. A browser tab cannot listen, cannot see which application is focused, and cannot type into it.