How does AI remove filler words from dictation?

A language model removes filler words by reading the transcript and deciding which words add nothing. The usual pipeline has two steps. A speech recognizer turns your voice into text, and then a language model edits that text. The model sees words, not sound, so the recognizer's output decides what can be removed. Fillers the recognizer dropped cannot be removed, and words the model reads as clutter can be deleted even when they carry meaning. The steps and the design choices are set out in our guide to AI cleanup for dictation.

The model works on the transcript, so the recognizer comes first

In the usual design, the recognizer writes out a raw transcript, and the cleanup step then edits that transcript. Microsoft says that Fluid dictation, a Voice Access feature on Copilot+ PCs, automatically corrects grammar, punctuation and filler words as you speak, which reduces the need for manual editing ¹. Spokenly describes an optional on-device cleanup model that removes filler words and repeats, applies corrections you say out loud, fixes punctuation and capital letters, and writes numbers as digits ². Some products may combine these steps, so check the vendor's documentation for the one you use.

This ordering has a consequence. If the recognizer leaves out a filler, the model has nothing to delete. A 2025 preprint on dementia speech notes, as a side observation, that Whisper tends to omit filler words such as uh and um from its transcripts ³. Many ASR systems discard or generalize hesitations, losing important acoustic detail ⁴. So two apps can produce very different results from the same speech. One may have fillers to remove, and the other may have none.

Fillers are removed as text, which is why deletion is low risk

Deleting a filler is the easiest edit a model makes, because a filler carries no information. Readability studies cited by Lease and Johnson in 2006 found that fillers and speech repairs can be deleted from transcripts without compromising meaning ⁵. The finding is about transcripts, not dictated drafts, and it is old, but it fits how cleanup works when the rest of the sentence stays put.

Fillers also come from somewhere. In a corpus of conversations, disfluency rates were higher when speakers discussed abstract figures, which the authors read as an increase in planning difficulty ⁶. In conversation, then, a filler can mark the moment a speaker was still deciding what to say. Removing it fixes the page, not the planning. For ways to reduce the fillers you produce, see how to stop saying um when you dictate.

The model removes four kinds of material, and they are not equally safe.

Hedges are the trap. A hedge is not a filler, but a model that reads it as clutter can delete it. "We will probably ship on Friday" can come back as "We will ship on Friday," which is a different claim. The same risk applies to words like "like" and "so". "I like it" must keep its "like", and "like, you know" is a filler.

Good cleanup depends on what the model is allowed to change

The model can remove fillers well when the instruction is narrow. Give it a prompt that names the fillers, asks it to apply only corrections you clearly made, and forbids rewording. Without that limit, a model that is asked to "clean up" may also rephrase and reorder, and then the fillers are not the only thing it removes. A 2023 evaluation of ChatGPT as a grammar corrector found that its very fluent corrections were possibly due to over-correction and not following the principle of minimal edits ⁷. Minimal edits are what you want for filler removal.

The glossary entry on filler words defines the terms used here. For the wider question of which edits are safe, see does AI dictation change what you said?.

What to do to remove fillers without losing meaning

1. Choose a tool that removes fillers on purpose. If the feature is a setting, turn it on. If it is a prompt, use one that names the fillers.
2. Ask for minimal edits. Say which items to remove, which corrections to apply, and that nothing else should change. The prompt in how to clean up dictation with ChatGPT is one example.
3. Keep the raw text. You need it to check the result and to undo a bad deletion.
4. Check the hedges, negatives and numbers against what you said, since these are the words a model most often deletes or changes by mistake.
5. Use the undo route when it exists. In Windows Fluid dictation, you can say "Revert" or "Undo that", or press Ctrl+Z, to keep the original sentence ⁸.

Once filler removal works in your tests, move on to rewording, which carries more risk. The guide to AI cleanup explains the five cleanup levels, and the level that removes fillers and fixes punctuation without rewording is the place to start.