Fundamentals

How to Remove Filler Words From Dictation: Um, Uh and False Starts

To remove filler words from dictation, you need a cleanup step after speech recognition, because an accurate transcriber writes down your um, uh and false starts just as you said them. AirTypes Polished runs that cleanup before the text reaches your cursor, so the fillers never land in your document. Plain is free and offline and removes only a leading um; Polished cleans up the whole sentence at 120 minutes for every dollar.

Here is a real example of how people actually talk when they explain something, and exactly what Polished typed:

You say

"so um I think we should uh probably move the deploy to friday because like the migration isn't, the migration isn't tested yet and uh QA needs two more days"

→
Polished types

"I think we should move the deploy to Friday because the migration isn't tested yet and QA needs two more days."

Why speech recognition keeps filler words

Nobody speaks in finished sentences. While you work out what to say next, you fill the gap: "um", "uh", "so", "like". You start a phrase, notice it is going the wrong way, and start it again. Listeners filter all of this out without noticing. A page does not.

A speech recognizer is measured on one thing: how closely its text matches the words in the audio. From that point of view, "um" is a word you said, and "the migration isn't, the migration isn't" is two phrases you said. Writing them down is correct behavior for a transcriber. It is just not what you want in an email.

Models like OpenAI Whisper were trained on large amounts of audio paired with text, much of it already tidied by people, so they drop some fillers on their own. Which ones they drop depends on the recording and the model, and it is not something you can rely on. You might get a clean sentence one time and an "uh" in the middle of the next. False starts and repeated words are even harder, because they are real words in the right order; nothing in the sound says "ignore this part". If you want to know more about how the models differ, see Whisper model sizes explained.

So removing filler words is a separate step. It needs something that reads the transcript as a sentence and decides which words carry meaning.

Not all filler is the same

"Filler" covers several different habits, and they are not equally easy to remove:

  • Hesitation sounds: um, uh, er, hmm. Almost never meaningful, easy to spot.
  • Discourse markers: so, ok so, like, you know, I mean, basically. Often filler, but "like" in "I like it" or "looks like rain" is a real word, and so is "so" in "so that it retries".
  • Hedges: probably, I think, kind of. Sometimes padding, sometimes the most important word in the sentence.
  • False starts and repeats: "can we, can we talk". Only one copy should survive.
  • Self-corrections: "send it Tuesday, no, Wednesday". The first version has to go, and only an understanding of the sentence tells you which one.

A word list handles the first group. Everything after that needs judgment about meaning, which is why a simple rule never quite gets dictation clean.

What Plain does: the leading um only

AirTypes has two modes, switched on the Home screen. Plain runs Whisper offline on your own CPU, uploads 0 bytes of audio, works in airplane mode and is English only. After transcription it applies a light cleanup on your machine: it removes a leading um or uh, capitalizes the first letter and adds a period. That is all. Here is a real Plain result:

You say

"ok so um the meeting is moving to tomorrow at ten, client is not available, and uh send him the invoice again"

→
Plain types

"The meeting is moving to tomorrow at ten, client is not available, and uh send him the invoice again."

The opening "ok so um" is gone, the first letter is a capital and there is a period at the end. The "uh" in the middle is still there, and so is the missing "the" before client. Plain is deliberately conservative: it never rewrites your sentence, so it cannot change what you meant. For short notes, search boxes, chat messages and anything you will edit anyway, that is often enough. It costs nothing and nothing leaves your computer.

What Polished does

Polished sends the recording to AirTypes to transcribe and clean it up. A large speech model on our servers transcribes it, then an AI removes filler words and false starts and fixes grammar and punctuation, keeping your meaning, numbers and names. Here is the same sentence Plain typed above:

You say

"ok so um the meeting is moving to tomorrow at ten, client is not available, and uh send him the invoice again"

→
Polished types

"The meeting is moving to tomorrow at 10:00. The client is not available. Send him the invoice again."

Every filler is gone, including the "uh" in the middle. The run-on became three sentences, "client" got its article, and the time became 10:00. Now look back at the deploy example at the top of this post. "so um", "uh" and "like" were removed, and the false start "the migration isn't, the migration isn't" became one clean clause.

Be aware that it also dropped "probably". In that sentence it read as padding next to "I think", and most readers would agree. But if a hedge matters to you, for example when you are deliberately not committing to a date, read the result and put it back.

A third real example, a message to a colleague:

You say

"hey Priya so uh I looked at the draft and I think section two is good but um section three needs more data, can we, can we talk thursday at four"

→
Polished types

"I looked at the draft. Section two is good, but section three needs more data. Can we talk Thursday at four?"

The repeated "can we" collapsed into one, a question mark appeared, and Thursday got its capital. The greeting "hey Priya" was dropped too. Polished can treat a greeting as filler, so if you want one, type it first and dictate the rest. Long-form writing gets the same treatment. In another real example, "the results suggest that um the effect is, the effect is stronger in the treatment group, which is consistent with uh Smith twenty nineteen" came out as "The results suggest that the effect is stronger in the treatment group, which is consistent with Smith 2019." More on that workflow in our guide to voice dictation for academics.

Talk the way you think. Type without the um.

Plain is free and offline. Polished removes fillers and false starts as it types, on prepaid credit, 120 minutes for every dollar.

Download AirTypes

The guard: when you get your own words back

Any cleanup that understands meaning can also misunderstand it. A tool that confidently deletes the "no" in "send it Tuesday, no, Wednesday" has made your message worse, not cleaner.

Polished has a guard for this. It checks that the rewrite keeps every number you said and stays close to your length. If not, AirTypes does not type the rewrite; you get your own words, the raw transcript, instead. We see this most with requests that ask an AI for code. In one real test, "ok so write me a um python function that takes a list of dates and uh returns the ones that fall on a weekend, and add tests, use pytest not unittest" came back unchanged, fillers included, because the rewrite drifted from the request. An "um" left in is a small cost. A changed instruction would be a bigger one.

When the guard kicks in, either delete the stray fillers by hand or say the sentence again a little more cleanly. Dictating prompts to chat assistants has a few habits of its own; voice to AI prompt covers them.

Manual ways to remove filler words

If your dictation tool does not clean up for you, there are three common workarounds. Each one works, and each has a cost.

1. Find and replace

Most editors support regular expressions in find-and-replace. A pattern like \b(um|uh|er|erm)\b,?\s* replaced with nothing removes hesitation sounds. It is fast and private. It also leaves broken capitalization when the filler started a sentence, double spaces and stray commas, and it does nothing for false starts or repeats. Do not add "like", "so" or "I mean" to the pattern; you will delete real words.

2. Ask a chatbot afterwards

Paste the transcript into ChatGPT, Claude or Gemini with an instruction such as "Remove filler words and false starts. Fix punctuation. Do not change the wording otherwise." This handles all five kinds of filler well. The costs: you copy, switch windows, paste, wait, copy again, switch back and paste, for every piece of text. Chatbots also tend to rewrite more than you asked, so you read the result carefully anyway. And the text goes to that chatbot's provider.

3. Edit by hand, or speak more carefully

You can read every dictated sentence and fix it, or train yourself to pause silently instead of saying "um". Both work. Both cost the speed that made dictation attractive in the first place, and speaking carefully tends to make you lose the thread of what you were saying.

Why cleaning at the cursor saves the round trip

The real cost of filler words is not the words. It is the extra steps between saying something and having clean text where you need it. Compare the two workflows for one email paragraph:

Dictate, then clean in a chatbotPolished at the cursor
Dictate into the emailClick into the email
Select and copy the textHold Ctrl + Shift + Space and speak
Switch to the chatbot and paste with an instructionLet go
Wait, then copy the replyRead it
Switch back, paste over the original, fix formatting

With AirTypes, the cleaned text is typed at your cursor in any desktop app: your mail client, Slack, a document, a code editor, a terminal. There is no plugin and no window to switch to, so you stay in the place you were writing. The hotkey can be changed, and you can choose hold-to-talk or toggle in Settings. For the full tour, see how to use AirTypes.

If your speech has grammar slips as well as fillers, the same step fixes those too; dictation for non-native English speakers walks through a real example. And if you think in another language, Polished works in the language you speak, with an optional Always translate to English setting: see voice typing Bangla to English.

Cost and privacy

Plain is free, with no time limit. Polished is pay as you go on prepaid credit, and your first minute is free. You recharge any whole-dollar amount from $5. $5 buys 600 minutes of speech and $10 buys 1,200, charged by the second with one price for transcription and cleanup together. A 30-second note costs under 1¢. Credit never expires, there is no subscription, and when credit runs out Polished pauses and Plain keeps working. Details are in the Polished section and on pricing.

Plain uploads 0 bytes of audio. Polished sends the recording to AirTypes to transcribe and clean it up. In both modes AirTypes receives your account email and usage counts. For the full picture, read is voice dictation private?

FAQ

How do I remove um and uh from a transcription?

For a transcript you already have, use find-and-replace with a pattern for um, uh and er, then fix the capitals and commas it leaves behind. For new dictation, AirTypes Polished removes fillers and false starts before the text is typed, so there is nothing to clean afterwards.

Why does speech to text keep my filler words?

A speech recognizer's job is to write down the words in the audio, and um, uh and false starts are words you said. Some models drop a few fillers on their own, but not reliably. Removing them is a separate cleanup step.

Does the free Plain mode remove filler words?

Only a leading um or uh. Plain also capitalizes the first letter and adds a period, all offline on your CPU. Fillers in the middle of a sentence and false starts stay as you said them.

Can removing filler words change my meaning?

It can if the cleanup misreads the sentence. Polished is told to keep your meaning, numbers and names, and a guard types your own words instead when a rewrite drops a number or drifts too far from what you said. It may drop a hedge like "probably" or a greeting, so read anything important.

Does filler removal work in languages other than English?

Yes, in Polished. It works in the language you speak, including Hindi, Bangla, Japanese, German and most other spoken languages, so try yours. Plain is English only.

Get AirTypes, free

Plain is free and offline for English dictation in any app. Add Polished when you want fillers, false starts and grammar cleaned up as you type, on prepaid credit with no subscription.

Download for macOS and Linux