> ## Documentation Index
> Fetch the complete documentation index at: https://docs.freestylevoice.com/llms.txt
> Use this file to discover all available pages before exploring further.

# AI cleanup

> Turn raw, rambly speech into polished text with a language model.

Speech is messy. You backtrack, you say "um", you start a sentence three times before it lands. **Cleanup** runs your raw transcript through a language model that fixes it up before the text reaches your app — so what you *meant* to say is what gets typed.

It's optional and off until you pick a cleanup model. Turn it on under **Settings → Models**.

## What it does

Every cleanup level does the same core job: add punctuation and capitalization, drop filler words, and resolve the moments where you correct yourself mid-sentence. The difference is how far it goes beyond that.

<Note>
  When you retract something ("send it to marketing, actually no, to legal"), cleanup keeps only your final intent ("Send it to legal.") — the abandoned wording is dropped, not left in the text.
</Note>

## Cleanup intensity

Pick how aggressive the edit should be under **Settings → Models**.

| Level      | What it does                                                                                                                                                                  | Before → after                                                                                                                                                                                                |
| ---------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Low**    | Smallest possible edits. Punctuation, capitalization, filler removal, and self-corrections only. Keeps your exact wording, including casual shorthand like "gonna" and "cuz". | "um so we're gonna push the demo back a bit cuz we found some issues" → "We're gonna push the demo back a bit cuz we found some issues."                                                                      |
| **Medium** | Sentence-level grammar and phrasing fixes without changing meaning, order, or detail. Reads like you wrote it carefully.                                                      | "i was thinking like maybe we could uh push the the deadline to friday because the designs aren't ready" → "I was thinking maybe we could push the deadline to Friday, because the designs aren't ready yet." |
| **High**   | Rephrases and reorders for flow and concision while preserving every fact, instruction, and caveat.                                                                           | "so there's like three things we gotta do um update the docs and uh notify support oh and the server needs a restart" → a clean numbered list of the three tasks.                                             |
| **Custom** | Write your own instructions and control exactly how text is rewritten.                                                                                                        | Whatever you tell it to do.                                                                                                                                                                                   |

<Warning>
  No level translates your text or invents facts. Cleanup always keeps the same language and script you spoke in, and never adds content you didn't say.
</Warning>

### List detection

At every level, when you dictate a sequence ("first… second… then…" or "one… two… three…"), cleanup formats it as a proper list instead of a run-on sentence. Ordered steps become numbered lists; plain item lists become bullets.

## Custom prompts

Need something specific — always draft in the third person, always output Markdown, always keep a certain tone? Choose **Custom** intensity, hit **View / edit prompt**, and write your own instructions.

<Tip>
  Editing any built-in preset automatically switches you to **Custom** and seeds the editor with that preset's text, so you can tweak from a solid starting point without overwriting the built-ins.
</Tip>

## Where cleanup runs

You can run cleanup on any of these, chosen under **Settings → Models**:

* **Freestyle Transcribe** — transcription and cleanup in one fast pass.
* **A local LLM** — runs on your device, nothing leaves your computer.
* **A cloud model** — bring your own key for **OpenAI**, **Anthropic**, **Google**, **Groq**, or **Mistral**.

Cleanup also adapts to *where* you're typing: if a [format](/features/formats) matches the active app, its instructions steer the rewrite (for example, a more formal register for email). See [Formats](/features/formats) for details.

## How it fits the pipeline

Cleanup is one stage in a larger flow. Roughly:

<Steps>
  <Step title="Vocabulary biases recognition">
    Your [vocabulary](/features/vocabulary) terms are sent to the speech model so tricky names and jargon come through correctly.
  </Step>

  <Step title="Speech is transcribed">
    Your audio becomes raw text.
  </Step>

  <Step title="Cleanup rewrites the text">
    The language model tidies it up at your chosen intensity (this step).
  </Step>

  <Step title="Dictionary applies replacements">
    Your [dictionary](/features/dictionary) swaps exact phrases in the cleaned text. This runs even when cleanup is off.
  </Step>
</Steps>

<Note>
  If cleanup fails for any reason (a provider hiccup, an unsupported model), Freestyle falls back to your raw transcript rather than losing your words.
</Note>