Why your dictation sounds bad: filler words and how AI cleans them up

You speak a clean, sharp sentence. You hit the dictate button. Out comes:

"So, um, basically what I want to say is, like, you know, I think we should probably, uh, ship this on Friday or something."

Painful. Not because the speech recognition got anything wrong — it got everything right. Every "um," every "like," every "you know." That's the problem. The dictation is too faithful.

Why filler words are invisible when you talk

Linguists call these discourse markers or hesitation particles. They serve real social functions in conversation:

• "Um" and "uh" hold the floor while you think. Without them, the other person assumes you're done and starts talking.
• "Like" approximates — "it's, like, ten dollars" means "around ten, give or take." It softens claims.
• "You know" checks comprehension — "this part, you know, where the API rejects the request" cues the listener that they should recall context.
• "So" opens a thought or signals a conclusion. It connects the next sentence to what came before.

Your brain filters all of this out when you listen. Your brain does not filter it out when you read.

What straight transcription gets you

Most off-the-shelf speech-to-text — Windows' built-in Win+H, macOS Dictation, the recorder app on your phone — does verbatim transcription. That's the right call for some applications:

• Court reporting
• Subtitling video
• Transcribing interviews for journalism
• Anything where the speaker's voice is the content

But it's the wrong call for dictation that's going to be typed into a Slack message, an email, a code comment, a document. There, the speech is just an input method. The output should read like something you'd write.

The cleanup problem isn't just removal

A naive filter would strip "um" and "uh" and call it done. That misses most of what makes dictated text look raw:

1. Self-corrections: "I think we should — actually let me rephrase — I think we should ship it Friday." You meant the second version. The first half is mental scratch paper.
2. Restart fragments: "The thing is — what I'm saying is — we need to..." Three false starts before the real sentence.
3. Listy speech: "We need to do A and we need to do B and we need to do C." That should be a bulleted list, not a run-on sentence.
4. Missing punctuation: speech doesn't have commas or periods. You hear them; the dictation engine doesn't always.
5. Email/chat conventions: when you say "comma" out loud you don't usually mean it literally — you mean the listener should imagine one. Same for "new paragraph."

How AI cleanup actually works

Pithflow runs every transcript through a second pass — a small AI model tuned for cleanup. Its instructions, in plain English:

1. Remove discourse markers (um, uh, like, you know, so, basically).
2. Apply the user's most-recent self-correction as their intended phrasing.
3. Add punctuation as a reader would expect.
4. Don't paraphrase — keep the user's wording. Cleanup is mechanical, not creative.
5. Preserve language — don't translate Spanish input to English.
6. Apply the chosen tone — casual, professional, Slack, code comment, etc.

Step 4 matters. A model that paraphrases will smooth your voice into something generic and corporate. The point of dictation is YOUR words appear faster — not "your words run through a marketing intern."

Before and after — three real examples

1. Slack message

Raw transcript:

"Hey um so basically I was looking at the dashboard you know and like the numbers look off for last week. I think maybe we should you know double check the data pipeline or something."

Cleaned (Slack tone):

"Hey — I was looking at the dashboard. The numbers look off for last week. Should we double-check the data pipeline?"

2. Email

Raw transcript:

"Hi Sarah uh thanks for the feedback on the proposal. So I think um what we want to do is uh basically incorporate the changes you suggested on um page three and uh send it back by Friday like end of day."

Cleaned (email tone):

"Hi Sarah, thanks for the feedback on the proposal. I'll incorporate the changes you suggested on page three and send it back by end of day Friday."

3. Code comment

Raw transcript:

"Okay so this function um basically takes a user ID and you know returns the most recent transcription. Returns null if there are no transcriptions for that user."

Cleaned (code-comment tone):

"Takes a user ID and returns the most recent transcription. Returns null if there are no transcriptions for that user."

When you DON'T want cleanup

There are real cases where the raw transcript matters more than the cleaned-up version:

• Captions for video — the voiceover should match what was said.
• Transcribing an interview verbatim — the hesitations carry meaning ("the suspect said 'um' for six seconds before answering").
• Notes about how you sound — coaching your own presentation style.

For these, turn cleanup off. In Pithflow that's a single toggle in the Style tab — and the raw transcript is always preserved in your local history, so even with cleanup on you can recover the verbatim version later.

The 80% rule

Most people don't write the way they speak. They speak in fragments, restart sentences, and repeat themselves. They also know — when they read what they spoke — that this is not how I write.

Filler-word cleanup closes ~80% of that gap automatically. The remaining 20% is intentional wording choices, structural edits, and the things you actually need to think about. Cleanup doesn't replace your judgment. It replaces the boring mechanical pass that you'd do otherwise — strip the "ums," capitalize sentences, add periods, fix the word salad — so by the time the text lands on your screen, you can focus on the parts that actually need a brain.

---

Pithflow is voice dictation for Windows that runs AI cleanup on every transcript. Download free and try it. The free tier gets 2,000 words/week — plenty to feel whether AI cleanup is worth it for you.