# Disfluency Canonical page: https://flit.fyi/glossary/disfluency > A disfluency is any break in fluent speech: fillers like "um" and "uh", repeated words, false starts, and self-corrections. Natural speech is full of them; finished writing has none. [Glossary](https://flit.fyi/glossary) / [What comes out](https://flit.fyi/glossary#output) Also: fillers Updated October 4, 2026 Definition A disfluency is any break in fluent speech: fillers like "um" and "uh", repeated words, false starts, and self-corrections. Natural speech is full of them; finished writing has none. Removing disfluencies is the first and most visible part of turning a transcript into writing. The subtle part is telling filler from meaning: a repeated word can be a stumble or emphasis, and a false start can be noise or a deliberate correction ("three hundred, scratch that, four hundred dollars"). Tools handle this anywhere from not at all (raw transcripts) to rule-based filler removal to a full language-model cleanup pass, with the risk at the far end being rewording things you meant. ## The kinds of disfluency Filled pauses: um, uh, er, and their equivalents in every language. Discourse markers: like, you know, sort of, when they carry no meaning. Repetitions: "the the meeting", "I I think". False starts: a phrase abandoned and restarted. Self-corrections: a value replaced by another, signaled by "no wait", "scratch that", "I mean". Each needs different handling, and the last one is the interesting case, because the right output is not the removal of words but the survival of the correction. ## Rules, models, or both Filled pauses and repetitions are safe to remove by rule: "um" is never a word you meant. Discourse markers are trickier, because "like" is sometimes a verb. Self-corrections are the strongest case for a narrow rule: when both sides of "no wait" are the same kind of thing, a time or a number or a day, replacing the first with the second is certain, and a model is not needed. Everything looser belongs to a language model with a strict brief: remove what was not meant, never rewrite what was. ## The failure mode to watch for Overcleaning. A model told to make speech read well will, given the chance, tighten your sentences, swap your words for its own, and drop the hedge that was the whole point. The test is simple: dictate something with an unusual word choice and see whether it survives. A tool that removes your fillers and keeps your words is doing the job; one that improves your prose is doing a different job you did not ask for. ## How to dictate with fewer fillers The cleanest way to deal with fillers is to say fewer of them. A filler is usually a pause filled with sound, and the habit comes from conversation, where silence invites someone else to start talking. A dictation tool does not interrupt, so you can pause in silence instead. With Apple’s built-in Dictation you have plenty of room to do that. Apple states that Dictation stops automatically only when no speech is detected for 30 seconds. 1. Decide what the sentence is before you start the dictation, even if only roughly. 2. Speak one sentence at a time and stop at the end of it. 3. When you need to think, close your mouth and say nothing. A silent pause leaves nothing in the text. 4. Slow down slightly. Fillers appear when your mouth runs ahead of the next thought. 5. If you make a mistake, finish the sentence and correct it afterward instead of restarting it halfway. 6. Read back what you dictated and note which fillers you use most, so you can listen for them next time. ## Why is "um" not counted in accuracy benchmarks? Because many benchmark setups delete fillers from both transcripts before they count errors. The text normalizer published with the Whisper paper removes the words hmm, mm, mhm, mmm, uh, and um from the reference and from the model output before word error rate is calculated. The practical result is that a published accuracy figure tells you nothing about fillers. A model that writes every um and a model that writes none of them can receive the same score. Whether fillers reach your document is a property of the tool around the model, and you have to test it yourself. ## How to test how a dictation tool handles disfluencies A short test separates the tools that remove fillers by careful rule from the ones that remove too much or nothing at all. Dictate each line below as written, hesitations included, and compare the output with what you meant. 1. Dictate a sentence with filled pauses, such as: "the report is um almost uh finished". Both fillers should be gone and every other word should remain. 2. Dictate a stumble, such as: "I I think the the meeting moved". The doubled words should appear once. 3. Dictate a deliberate repeat, such as: "that was a very very long day". Both words should survive, because the repeat is emphasis. 4. Dictate "like" used as a real word, such as: "I like the second option". The word must not be removed. 5. Dictate a self-correction, such as: "send it Monday no wait Tuesday". Check whether you get the corrected value, both values, or the wrong one. 6. Dictate a hedge you mean, such as: "I think we should probably wait". If "I think" or "probably" is missing, the tool is rewriting, not cleaning. In Flit Flit removes fillers and false starts, honors "scratch that" and "no wait" by rule, and throws away any rewrite that loses a word you said. [How Flit formats](https://flit.fyi/#writes) - [“the budget is uh three hundred dollars scratch that four hundred”](https://flit.fyi/#demo) - [Correcting yourself, and the phrases Flit listens for](https://flit.fyi/commands#corrections) - [Why delivered WPM is lower than speaking rate](https://flit.fyi/blog/dictation-vs-typing) ## Questions people ask. ### Does dictation software remove ums and uhs? The good ones do, automatically, along with repeated words and false starts. Raw transcribers leave them in. Built-in dictation on most systems leaves most of them in too. ### What happens if I correct myself while dictating? With a tool that handles self-corrections, saying "Monday, no wait, Tuesday" produces Tuesday. The safest implementations do this by rule when both values are the same kind of thing, and leave looser cases to the formatting model. ### Can filler removal change what I meant? Rule-based removal of ums and repeats cannot. A free-rewriting language model can, by tightening or rewording. Look for a tool that promises your content words survive, and test it with a sentence you would notice being changed. ### How do I stop saying um when I dictate? Replace the sound with silence. Decide the sentence before you start, speak one sentence at a time, and pause without speaking when you need to think. A silent pause adds nothing to the text. ### Do filler words count as errors in word error rate? Often they do not. The Whisper paper’s text normalizer, for example, removes hmm, mm, mhm, mmm, uh, and um from both transcripts before scoring, so fillers neither help nor hurt the published number. ### What is the difference between a filler and a false start? A filler is a sound or word that fills a pause, such as um or uh. A false start is a phrase you begin, abandon, and say again differently. Both are disfluencies, but a false start contains real words, so removing it takes more care. ## Guides that use it - [Mac dictation keeps stopping](https://flit.fyi/mac-dictation-keeps-stopping) - [ChatGPT dictation on a Mac](https://flit.fyi/dictate-to-chatgpt-mac) ## On the blog - [The 15 best Wispr Flow alternatives in 2026](https://flit.fyi/blog/best-wispr-flow-alternatives) - [Dictation vs typing: what the research says](https://flit.fyi/blog/dictation-vs-typing) ## Use cases - [Writers](https://flit.fyi/use-cases/writers) - [Students](https://flit.fyi/use-cases/students) ## Compared on it - [Flit vs Apple Dictation](https://flit.fyi/alternatives/apple-dictation) ## Related terms - [Formatting pass](https://flit.fyi/glossary/formatting-pass) - [Verbatim mode](https://flit.fyi/glossary/verbatim-mode) - [Words per minute](https://flit.fyi/glossary/words-per-minute) - [Dictation](https://flit.fyi/glossary/dictation) ## Your voice was always faster. 7 days free, then $19 USD once. No account, no subscription, nothing uploaded. [Download for Mac](https://flit.fyi/download) [Buy now · $19 USD](https://flit.fyi/pricing#personal) macOS 14 or later · Apple silicon