AI Cleanup for Dictation: Fillers, Punctuation, Rewrites

This guide belongs to our voice and AI overview. AI cleanup is a second pass that a language model makes over dictated text after the speech recognizer has written it down. A light pass removes ums, uh, "you know" and false starts, applies the corrections you speak aloud, and adds punctuation and paragraph breaks you never said. Heavier passes rewrite for tone or flow, and that is where the risk sits: the same model can drop a qualifier, change what you meant or add a sentence you never spoke. This guide sets out five cleanup levels from verbatim to full rewrite, with examples, the research on each kind of edit, and a prompt that keeps the edit faithful.

AI cleanup turns a spoken draft into written text in one pass

Dictation has two stages. A speech recognizer turns sound into words, and then an optional language model edits those words. The recognizer hears you. The cleanup pass decides what the page should look like. For the whole pipeline and the models behind it, see AI dictation: how AI voice typing works.

Cleanup is now a built-in feature in several products. Microsoft says that Fluid dictation, a Voice Access feature on Copilot+ PCs, automatically corrects grammar, punctuation and filler words as you speak ¹. On iPhone with Apple Intelligence in iOS 27, dictation uses an on-device model to improve spelling, punctuation and capitalization, and Apple lists that feature as available in English on named iPhone models ². Third-party apps offer it as well. Spokenly's optional cleanup model removes filler words and repeats, applies corrections you say out loud, fixes punctuation and capital letters, and writes numbers as digits ³.

Those features bundle five separate jobs: removing fillers, removing repeats and false starts, applying self-corrections, adding punctuation and layout, and rewording. The first four edit the speech you dictated. The fifth writes new sentences. Most of the risk comes from the fifth, so this guide treats each job on its own. Before you pick a tool, decide which jobs you want. The apps that offer them are compared in AI dictation apps compared.

Fillers, false starts and self-corrections are the easiest wins

Removing filler words is among the lowest-risk edits a language model makes. Readability studies cited by Lease and Johnson in 2006 found that fillers and speech repairs can be deleted from transcripts without compromising meaning ⁴. That is an old finding, but it still fits how cleanup works: a filler carries no information, so removing it is low risk when the rest of the sentence stays put. The question of how the model finds fillers is answered in how does AI remove filler words from dictation?.

Three kinds of material need removing, and they are not equally easy to spot:

• Fillers are sounds and phrases with no content, such as "um", "uh" and "you know". See the glossary entry on filler words.
• False starts are sentences you abandon and restart, such as "we should, we need to meet".
• Self-corrections are changes you make out loud, such as "Tuesday, no, Thursday".

Whether a filler reaches the cleanup pass at all depends on the recognizer. A 2025 study of dementia speech found that Whisper tends to omit filler words such as uh and um from its transcripts ⁵. Other engines keep them. Many ASR systems discard or generalize hesitations, which loses acoustic detail that verbatim transcription depends on ⁶. So a cleanup pass may have nothing to remove on one engine and a great deal to remove on another.

Keeping hesitation detail in the recognizer's output can also improve the accuracy of the text. In a 2025 challenge on non-native English, a fine-tuned Whisper system scored 5.81 percent word error rate when it labeled hesitations, against 6.47 percent when it removed them ⁷. The word error rate figure is a standard way to score a transcript, and the chart below shows the gap.

Self-corrections are harder than fillers, because the model has to decide which words the speaker wanted to keep. "Let's meet Tuesday, no, Thursday" should become "Let's meet Thursday." But "no" is also a word people say to mean no, as in "No, the budget is fine." A good prompt tells the model to apply a correction only when the speaker signals one clearly, and to leave anything uncertain in place. Windows also gives you a way back after the fact: to keep the original sentence or undo a Fluid dictation correction, you say "Revert" or "Undo that", or press Ctrl+Z ⁸. The Fluid dictation guide for Copilot+ PCs covers where the feature is available.

The practical point is that cleanup can only remove what the recognizer left in. It can add words back in some cases, as the next section shows, but it cannot reliably tell which dropped words you meant.

Language models also repair some misheard words from context

A language model can do more than delete. Given the recognizer's candidate transcripts, a large language model can choose a better word, and it can also produce a word the candidates never contained. Chen and colleagues report that in the HyPoradise benchmark, the model can correct words that are missing from the recognizer's N-best list, because it generates text rather than only choosing among candidates ⁹. The approach is called generative error correction, and the speech to text and LLMs guide explains the models in more depth.

Yang and colleagues report the same direction on out-of-domain tasks. Combining prompting with fine-tuning brought error rates below the N-best oracle level ¹⁰. Read that with care. These are research setups with labeled pairs of recognizer output and correct text, tested with 2023 models. They do not measure the dictation app you use today.

The everyday version is easy to picture. If the recognizer writes "the write email" for "the right email", a model that reads the surrounding words can pick the spelling that fits. The same ability explains the failure, too: a model free to write new words can write the wrong ones. The evidence for that comes later in this guide.

Adding punctuation is safer than rewording

Punctuation is the job most people ask for first. Szaszak's 2019 work on punctuation for speech recognition argues that automatic punctuation should favor precision over recall, because readers can add a missing mark in their heads more easily than remove a wrong one ¹¹. That is a sound rule for a cleanup pass: add the marks the model is sure about and leave the rest. If you want full control, speak the marks yourself, as covered in dictation punctuation. The setting that adds marks without a spoken command is described in the automatic punctuation entry.

Cleanup levels run from verbatim to full rewrite, and each one loses something

Cleanup is a dial, not a switch. Each step further from the spoken words saves reading effort and loses part of the record of what was said. The table uses one invented spoken sentence to show the five levels.

Spoken input: "um so we should, uh, meet Tuesday, no, Thursday"

Professional transcription uses the same ladder. Rev defines verbatim transcription as one that types every word heard, including filler words, false starts, self-corrections and stutters ¹². GoTranscript's clean verbatim keeps the meaning while removing obvious filler words, stumbles and minor repetition ¹³. Its edited transcription, at the far end, rewrites for clarity and flow. That helps readability but moves the text further from the spoken record ¹⁴.

For dictation into your own documents, level 3 is the sensible default. It removes the noise, applies the corrections you clearly made and fixes punctuation, while keeping your sentences. Level 4 suits a first draft you will read closely anyway. Level 5 produces new text, so treat it as a draft written by someone else. Use levels 1 and 2 when the transcript is a record, such as an interview or a clinical or legal note. The verbatim vs clean verbatim guide covers those formats.

You can also remove the language model entirely. Superwhisper's Voice to Text mode outputs raw transcribed text, with no language model involved ¹⁵, which gives you the level 1 result without any AI pass. The speech to text without AI guide covers that route in full.

Heavy rewriting can change meaning, so review what matters

The more a language model rewrites, the more likely it is to change what you meant. The research finds that the change can happen even when the model is told to do less. In a 2026 preprint, the authors found that LLMs not only alter the voice and tone of human writing but also consistently alter the intended meaning ¹⁶. Their most direct test asked the models for grammar edits only. Even then, the models still changed the text in a way that significantly alters its semantic meaning ¹⁷. A cleanup instruction that says "fix grammar" is not a guarantee of a faithful edit. The question of whether dictation changes your words is answered in more detail in does AI dictation change what you said?.

Cleanup can fail in two directions. The first is over-correction. A 2023 evaluation of ChatGPT as a grammar corrector suggested that its very fluent corrections were possibly due to over-correction, and to not following the principle of minimal edits ¹⁸. The second is over-deletion. A 2026 preprint on conversational speech found that reasoning models systematically over-delete fluent content, which the authors read as a bias toward abstraction over fidelity to the transcript ¹⁹. In practice the second failure looks like this (an invented example): "we will probably ship on Friday" becomes "we will ship on Friday" because the model treats "probably" as clutter. The meaning has changed, and nothing on the page shows it.

Vendors say the same thing about their own products. OpenWhispr's documentation describes its cleanup as conservative: it fixes obvious grammatical slips and otherwise leaves your wording alone ²⁰. The same documentation says that its default cleanup means that, occasionally, the text is not quite what you said ²¹.

The errors can be counted. In a UC Davis pilot of AI-generated clinical notes made from ambient recordings of visits, the most frequent error was an accidental omission ²². Hallucinations, meaning content the recording did not support, came next ²³, followed by accidental inclusions, meaning content from the visit that does not belong in the note ²⁴. Clinical notes are not dictated text, and the task is different. The change is the same kind, though: a language model turns recorded speech into written text, and the errors that appear are omissions and added content.

Cleanup can also flatten your voice

Cleanup can change how you sound even when each sentence is correct. In a Science Advances experiment, writers who received story ideas from a large language model produced stories that were rated as more creative, better written and more enjoyable, with the largest gains among less creative writers ²⁵. The same study found that stories written with generative AI ideas were more similar to one another than stories written by humans alone ²⁶. The authors call this a social dilemma: each writer gains, while the range of what gets written narrows.

Models also leave a vocabulary trace on the text they touch. An analysis of PubMed abstracts found an abrupt rise in certain style words after LLMs appeared, and it suggests that at least 13.5 percent of 2024 abstracts were processed with LLMs ²⁷. Opinion can move as well. In a 2023 study of co-writing with an opinionated language model, the model's suggestions affected the opinions expressed in people's writing and shifted their opinions in a later attitude survey ²⁸. For the direct question, see does AI cleanup make your writing sound generic?.

What to do about it: keep your hedges, your opinions and your sentence length. Look hardest at the words a model is most likely to rewrite on its own, such as adjectives, hedges and the first and last sentence of each paragraph.

Check an edit in five steps

1. Open the raw transcript next to the cleaned text.
2. Check every number, name, date and time against what you said.
3. Check every negative ("not", "never", "don't") and every hedge ("probably", "may", "I think"). These carry the meaning, so a changed one matters more than a changed comma.
4. Look for sentences you do not recognize. Those are additions.
5. Where accuracy matters, paste the raw transcript instead, or stay at level 2.

You can run the cleanup yourself with any chatbot and a fixed prompt

You can test cleanup without installing anything. Dictate into a chatbot, then paste a fixed prompt that limits the edit to level 3. The guide to dictation in ChatGPT and Claude covers the apps in more detail, and the question on how to clean up dictation with ChatGPT walks through the same steps.

1. Open ChatGPT in the desktop or mobile app, and tap the microphone icon in the text box. Your words then appear as text ²⁹.
2. Know where the audio goes. OpenAI says the recorded audio is sent to its models to be transcribed ³⁰. Do not dictate confidential material into a cloud chat you have not checked.
3. Put the prompt below above your transcript, and send it.
4. Compare the reply with your raw transcript using the five-step check above before you use it.

Clean up the dictated transcript below. Remove filler words such as um, uh, you know and like, and remove false starts. When I correct myself out loud, such as with "no" or "scratch that", keep only the corrected words. Add punctuation and paragraph breaks. Do not add, remove or reword anything else. Keep every number, name, date, negative and hedge, such as "probably" or "I think". If you are unsure whether a word is a filler or a correction, leave it in. Return only the cleaned text.

The prompt sets level 3 on purpose. Each instruction removes or punctuates, and none asks the model to rephrase, so the rewriting that the research warns about has little room to happen. Ask for level 4 only for a first draft you will read closely.

If the audio itself is the concern, you can keep dictation on the computer. Yap, which we make, is an on-device dictation app for Mac and Windows. Read how Apple dictation works and is Apple dictation private? to see what the built-in options keep, and on-device speech recognition for the trade-offs. For what a cloud cleanup costs in subscriptions and processing, see AI dictation cost.