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

# AI cleanup

> AI cleanup runs a language model over a dictation to strip filler and fix grammar and punctuation, on its own Dictate and clean up shortcut. Covers cleanup strength, fixing misheard words, where the model runs, and the timeout.

Speak messy, paste polished. Press *Dictate and clean up* instead of your normal dictate shortcut, and the text that lands has the filler stripped and the grammar and punctuation fixed.

| Platform          | Shortcut                     |
| ----------------- | ---------------------------- |
| Windows and Linux | `Ctrl` + `Shift` + `Space`   |
| macOS             | `Option` + `Shift` + `Space` |

Cleanup is off by default. Turn it on in *Settings → Dictation → AI cleanup*. It is not hidden behind the Experimental features toggle, and while the cleanup toggle is off that shortcut does nothing at all.

Your normal dictate shortcut still pastes raw text, so you choose per recording. Cleanup applies to dictation only, never to the assistant chat.

<Warning>
  Cleanup adds a short delay after you speak, because the model runs after transcription. If it fails, times out, or is not configured, your raw text is pasted instead. You never lose your words.
</Warning>

## The two controls you will actually touch

**Cleanup strength** decides how much gets rewritten.

| Strength               | What it changes                                                                                                         |
| ---------------------- | ----------------------------------------------------------------------------------------------------------------------- |
| **Light**              | Spelling, capitalization, punctuation, and obvious filler. Your repetition, sentence order, and exact phrasing survive. |
| **Balanced** (default) | The above, plus collapsing repeated points and dropping false starts.                                                   |
| **Aggressive**         | The above, plus tightening rambling, merging fragments, and reordering sentences for clarity.                           |

All three keep every fact, name, number, date, and request, and all three write in your first-person voice.

**Fix misheard words** is off by default. It lets the model replace a word that is clearly wrong for the sentence, like a homophone or a near-miss pronunciation. It stays off by default because it is the one place where cleanup is allowed to guess at your intent. Turn it on if speech-to-text mangles your accent often.

## Tone

Pick a **Writing style** to change how the result sounds. The default is `None (cleanup only)`, which fixes the text without changing its voice. Six styles ship with the app and you can save your own: see [Writing styles](/docs/writing/writing-styles).

Strength and style are separate dials. Strength decides how much gets tidied away. Style decides how whatever survives sounds.

## Setup and edge cases

<AccordionGroup>
  <Accordion title="Where cleanup runs, and the fallback that means you can skip this">
    Use the **Where it runs** toggle.

    * **On my device**. A local model you have downloaded. No key, no network.
    * **Cloud provider**. Any OpenAI-compatible provider with your own API key. The same list the assistant uses is available here: see [Providers](/docs/assistant/providers).

    On an Apple Silicon Mac, **Apple Intelligence** also appears in the provider list. It runs fully on-device with no key and no network, and needs macOS Tahoe 26.0 or later with Apple Intelligence enabled in System Settings. If it is not available, cleanup falls back to pasting your raw text rather than failing loudly.

    If the provider you selected here is not usable, meaning no model chosen or a cloud provider with no key, cleanup quietly falls back to your [assistant's](/docs/assistant/providers) provider and model instead. That is deliberate. Once the assistant works, cleanup works, with nothing extra to set up.
  </Accordion>

  <Accordion title="Editing the cleanup prompt">
    SpeakoFlow ships one prompt, **Improve Transcriptions**, and it is selected out of the box. You can edit it, add your own, and switch between them. You cannot delete the last remaining prompt, because cleanup needs one to run.

    Your transcribed text is sent to the model automatically, so just write instructions. For example: *"Fix grammar and punctuation, then rewrite as a polished, professional message."*
  </Accordion>

  <Accordion title="AI cleanup timeout">
    *Default: 10 seconds. Options: 5, 10, 15, 20, 30, 60.*

    How long to wait before giving up and pasting your raw transcription instead.

    This is the whole budget, not just the model call. On the built-in local engine, loading the model from disk on first use counts against it, so allow more time if your first cleanup of a session keeps timing out.
  </Accordion>

  <Accordion title="When cleanup does not run">
    Cleanup fails soft every time. Your raw transcription is pasted and the dictation completes normally. That happens when:

    * no provider is configured, or the selected one has no model, and neither does the assistant fallback
    * a cloud provider is selected but has no API key
    * no cleanup prompt is selected, or the selected one is empty
    * the provider returns an error, an unreadable response, or nothing at all
    * the model runs past the timeout

    History records both versions of a cleaned dictation, so you can always compare the original transcription with the final text.
  </Accordion>
</AccordionGroup>

<Note>
  Cleanup and [Generate with Flow](/docs/writing/generate-with-flow) are separate paths on separate shortcuts. Cleanup never runs over Flow output, and Flow never triggers on the *Dictate and clean up* shortcut.
</Note>