# The words are wrong

Find out in five seconds whether Blabb misheard you or cleanup rewrote you, then fix the right half of the problem.

Two completely different things produce the same complaint. Either the speech model misheard you,
or it heard you correctly and the cleanup style changed the words afterwards. The fixes have
nothing in common, so find out which one it is before you change a setting.

## Ask this first

Open the **History** window and find the dictation. Each entry can show three things:

| View | What it is |
| --- | --- |
| **Original words** | The raw transcription, exactly as the speech model heard you |
| **Final text** | What was actually typed into your document |
| **What changed** | The difference between the two |

**What changed** answers the question on its own. If the wrong word is already there in **Original
words**, the speech model misheard you. If **Original words** is right and **Final text** is wrong,
cleanup changed it.

[The History window](/docs/get-started/history/) covers searching and filtering if the dictation
was a while ago.

## If the speech model misheard you

Work down this list. The first two are worth more than the rest combined.

### Your microphone and how loud you arrive

Microphone quality and placement decide more about accuracy than any setting in Blabb. A headset
microphone a few centimetres from your mouth beats a good microphone across the desk.

- [Microphone setup](/docs/accuracy/microphone-setup/) — which microphone, and where to put it
- [Windows microphone levels](/docs/accuracy/windows-microphone-levels/) — arriving loud enough to
  be heard, and not so loud that you clip

### Your room

Hard surfaces, an open-plan office, a fan, a window onto a road. The model hears all of it.
[Your room](/docs/accuracy/your-room/) covers what is worth doing something about.

### How you are speaking

This one surprises people: the models are trained on natural speech at natural speed, so
over-enunciating and slowing down deliberately make recognition *worse*, not better. Speak the way
you would to a colleague. [How to speak](/docs/accuracy/how-to-speak/) has the detail.

### Specific words it keeps getting wrong

Names, products, acronyms and the local jargon are not in any general model's vocabulary. Teach
them once in **Saved words** — "when you say this, use this spelling" — and Blabb applies them both
as a hint to the speech model and as a correction pass over the raw transcription before cleanup
runs. If cleanup then changes a taught term, the rewrite is discarded, so the taught spelling
survives either way. See [Saved words](/docs/cleanup-styles/saved-words/).

### Whole fields of vocabulary

If it is not one word but hundreds — drug names, anatomy, case citations, Latin — that is what the
medical and legal vocabularies are for. They apply to every cleanup style, including Original.
See [Medical and legal vocabulary](/docs/editions/overview/).

## If cleanup changed it

This is usually the cleanup style doing exactly what it was asked to do. **Clear**, **Professional**
and **Friendly** are all allowed to rewrite; that is the point of them.

Two things to try:

1. **Move down the ladder.** **Original** keeps exactly what you said and runs no cleanup model at
   all. **Fix typos** corrects spelling and grammar and nothing else. See
   [Cleanup styles](/docs/cleanup-styles/overview/).
2. **Match the style to the writing.** A style that suits a chat message is the wrong tool for a
   clinical note or a quoted paragraph. If the exact words matter — figures, doses, names, anything
   you are quoting — stay on Original or Fix typos for that work, and save the heavier styles for
   prose where a better sentence is genuinely welcome.

If cleanup produced something you object to rather than something merely wrong, report it. The
**Report AI content** button sits on the history entry itself.

## Next

- [When words come out wrong](/docs/accuracy/when-words-come-out-wrong/)
- [Cleanup changed something it should not have](/docs/common-issues/unexpected-rewrites/)