Filler words in dictated text and how to clean them up
· 6 min read
A speech recognizer writes down what you said, and what most people say is full of “um”, “you know” and “sort of”. Some of it gets dropped on the way in, but not reliably, so a dictated draft usually needs one cleanup pass. The fastest fix is partly technique (pause instead of filling) and partly a find and replace you run once over the whole draft rather than word by word.
Why filler survives the trip
Recognizers are built to be faithful. If you make a sound that maps to a word, you get the word. “Um” and “uh” are borderline cases that some systems discard and others transcribe, and which you get can change with the language, the microphone and the background noise. Treat any dropping you notice as a bonus and not as a feature.
Real words used as filler always survive, because there is nothing to distinguish them from the same words used deliberately. “Like”, “so”, “right”, “basically”, “actually”, “I mean”, “you know”: each one is a legitimate word in some sentence, and a recognizer has no way to tell your throat-clearing from your meaning.
That leaves you with two jobs: say fewer of them, and remove the rest in one pass.
Say fewer of them: pause instead of filling
Filler words exist to hold the floor in a conversation. Dictating to a Mac is not a conversation, and nobody is waiting to interrupt you, so the sound you need while you think is silence.
Three things make that easier:
- Know the sentence before you start it. Most “um” arrives mid-sentence because the sentence was launched before it was finished. Think the clause, then say it.
- Use a held key rather than a toggle. With hold-to-talk you release the key while you think and press it again when you know what comes next, so the thinking never reaches the microphone at all. With a toggle, the microphone is open the whole time you hesitate.
- Watch out for inferred punctuation. Dictation systems that add punctuation from your pauses will read a long thinking pause as a comma or a sentence break. That is a fair trade against a transcribed “um”, but it means your cleanup pass also involves joining sentences that should not have been split.
Recording yourself for two minutes and listening back is the single most effective thing here. Everyone has one or two signature fillers, and hearing your own is usually enough to reduce them.
Build your own list before you build a tool
Before automating anything, write down the words you actually say. Look at a page of your own dictated text and count. A typical list runs to six or eight entries, and it is personal: someone who says “right” every third sentence may never say “basically” at all.
Split the list in two, because they need different treatment.
- Safe to delete anywhere: “um”, “uh”, “erm”, “you know”, “I mean”, “sort of”, “kind of”.
- Only sometimes filler: “so”, “like”, “right”, “just”, “actually”, “basically”, “obviously”. These need eyes on them. Deleting every “like” from a paragraph that compares two things will damage it.
Find and replace, once per draft
The unglamorous answer, and for most people the right one. Nearly every writing app on the Mac has Edit, Find, Find and Replace (Command-F, then the replace field). Put the filler in the search field, leave the replace field empty, and use Replace All for the safe list only.
Two details that save a mess:
- Include the trailing space in the search term, so “um ” rather than “um”, or you will leave double spaces behind.
- Turn on whole-word matching where the app offers it. Without it, searching for “um” will hit “number”, “album” and “summary”.
For the sometimes-filler list, use Replace one at a time and read each hit. It takes a minute and it is the difference between a tidy draft and a mangled one.
One line in Terminal, over the clipboard
If you dictate all day, wire the cleanup into a command. Copy the draft, run this, and the cleaned version is on the clipboard ready to paste:
pbpaste | perl -pe 's/\b(?:um+|uh+|erm|you know|I mean|sort of|kind of)\b,?\s*//gi' | pbcopy
Everything in that line ships with macOS. pbpaste prints the clipboard, pbcopy puts text back on it, and perl does the substitution. Use perl rather than sed here: the version of sed that comes with macOS handles word boundaries differently and the same expression will not behave.
Add your own words inside the parentheses, separated by the vertical bar. Then put it in ~/.zshrc as a function so it has a name:
defluff() { pbpaste | perl -pe 's/\b(?:um+|uh+|erm|you know|I mean)\b,?\s*//gi' | pbcopy; }
Open a new Terminal window, copy some dictated text, type defluff, and paste. If you would rather not go near Terminal at all, Automator does the same job as a Quick Action: New Document, Quick Action, set it to receive text in any application, check “Output replaces selected text”, add a Run Shell Script action with the perl line in it, and save. It then appears in the Services menu of every app, and you can give it a keyboard shortcut under System Settings, Keyboard, Keyboard Shortcuts, Services.
Let software rewrite it, with your eyes open
On a Mac that supports Apple Intelligence and has it turned on, select a passage, Control-click it, and look for Writing Tools. Proofread tidies grammar and awkward phrasing; Rewrite restates the passage. Both will remove filler as a side effect of the rewriting, and both will also change words you chose on purpose, so read the result rather than accepting it.
Voice Keyboard Pro takes the same idea a step earlier: its optional Smart Rewrite cleans filler words and shapes the tone as the text arrives, so what lands at the cursor is already tidied. That saves the pass, and it means the first version you see is not exactly what you said. For a chat message or an email that is a good trade. For a quote, a legal note or anything where the exact words matter, leave the rewriting off and clean up by hand; case notes by voice makes the same argument at more length.
Questions
Should I edit as I dictate or afterwards? Afterwards. Stopping to fix a word breaks the thread of the sentence you were building, and the whole speed advantage of dictation comes from not stopping. Say the draft, then clean it.
Do filler words matter in a meeting transcript? Less than you would think for a transcript people will search, and a lot for one people will read. If the transcript is going out to others, run the cleanup pass and fix the sentence breaks. Transcribing a meeting on an iPhone covers getting the recording in the first place.
Is there a way to stop “um” reaching the page at all? Not with certainty, since it depends on the recognizer. Hold-to-talk gets you closest, because the microphone is only open while you are actually saying something you mean.
Will removing filler make my writing sound artificial? Spoken language and written language have different rhythms, and a transcript that keeps every “you know” reads worse, not more human. Keep the contractions and the plain words; drop the noises. Voice commands while dictating covers shaping the text as you go.