Skip to main content

Cleanup levels: Verbatim, Light, Standard

How much Amlin is allowed to tidy what you said, and what it will never do at any level.

J
Written by Junan Pang

Set the default in Dictation Settings → Style → Default cleanup level, and override it per app in App profiles.

Verbatim

Exact transcript, with punctuation from the speech engine only. For quotes, legal work, and anyone who'd rather not have an AI touch their words at all. Filler words stay in. List detection is off. Spoken addresses, sites, and emoji names stay exactly as you said them.

Light: the default

  • Removes filler words and false starts

  • Fixes punctuation and capitalization

  • Applies self-corrections: "Wednesday, actually Tuesday" becomes "Tuesday"

  • Repairs an obvious mishearing when the sentence makes the intended word unmistakable: "send it two me in Sarah" becomes "send it to me and Sarah". When it is not sure, the word stays exactly as you said it.

  • Collapses stutters and repeated words

  • Formats numbers, dates, and years sensibly: "twenty twenty four" becomes "2024"

  • Turns a clear spoken enumeration into a numbered list

  • Writes spoken addresses and sites properly: "junan pang at gmail dot com" becomes [email protected], and "acme dot com" becomes acme.com. Say "underscore", "hyphen", or "plus" when you need one of those characters.

  • Types emojis you name: say an emoji's name followed by the word "emoji", so "thumbs up emoji" becomes 👍 and "rocket emoji" becomes 🚀

  • Capitalizes only when you are actually starting a sentence: dictate into the middle of a line you are already writing and the first word stays lowercase

Never reorders your ideas, substitutes your vocabulary, or changes your register.

Standard

Everything in Light, plus:

  • Sentence-boundary repair for rambling speech

  • Paragraph breaks in long dictations

  • Per-app writing style (casual / formal / excited) if you've set one

Standard's deeper polish comes from an optional on-device AI model. It's a one-time download: turn on Use the on-device cleanup model in Dictation Settings → Style. When it's installed, Amlin gives it a time budget that grows with how much you said, roughly a second and a half for a sentence and up to four for a long ramble. If the model finishes inside the budget, the refined text is what gets inserted. If not, the instant cleanup is inserted and the late result is discarded. Your text never waits beyond the budget, and Standard still applies all of its rule-based structure without the download.

Still forbidden at every level: paraphrasing, changing your tone beyond a style you explicitly chose, and adding content you didn't say. And your dictation is never answered: dictate a question and you get the question typed out, not an answer.

Spoken formatting

At all levels, spoken commands are honored: "new line", "new paragraph", "bullet point" / "next bullet", "quote … end quote". When it's ambiguous, Amlin favors the literal text.

Automatic lists

At Light and Standard, clear enumerations become lists. Number your points out loud, either "One, ... Two, ..." or "First, ... Second, ...", and each one becomes its own numbered line. Items that are questions keep their question marks.

Amlin is deliberately conservative here, because a wrong bulleting is worse than a missed one. Counting out loud, like "one, two, three", stays as you said it. Parallel adjectives inside a sentence stay inline. And the numbering has to run in order from the start, so "One, ... Three, ..." stays as prose.

Did this answer your question?