Turn voice notes into clean text with punctuation

You get clean text from a voice note when Parlo transcribes it with Apple's on-device speech engine, then cleans it up with an on-device language model. Cleanup removes filler sounds, false starts, and repeated words, and applies spoken commands like new paragraph. A length guard blocks any cleanup that adds words you did not say.

By Eric Friedman. Last updated 2026-09-08.

What happens between speaking and clean text?

Parlo runs two steps on the device, one after the other. First, Apple's SpeechAnalyzer and SpeechTranscriber turn your speech into a transcript. Second, an on-device language model cleans that transcript: it removes filler sounds, false starts, and repeated words, and may lightly fix punctuation and capitalization. Both steps run on your iPhone or Mac. Nothing about your voice note is sent to a cloud transcription service.

The result lands in your feed as one note. If you edit it later by hand or with a transform, Parlo keeps your dictated words on hand. A change you do not like is never the last word on your note (more on that below).

What does cleanup actually remove?

Cleanup targets the parts of speech that read badly on the page but say nothing:

Cleanup does not summarize, rewrite your meaning, or add new information. It cleans the sentence you already spoke.

What if cleanup changes too much?

A length guard checks the cleaned text against what you said. If the cleaned version comes back implausibly longer than the transcript, Parlo treats that as a sign the model added words you never spoke. It falls back to the deterministic version instead: filler sounds stripped, spoken commands applied, punctuation left plain. You always get a usable note, and it never grows content you did not say.

Can you turn cleanup off?

"Use AI cleanup" is a toggle in Settings. It is on by default on iPhone. Parlo for Mac ships with it off, and you turn it on there if you want it. With cleanup off, Parlo still strips filler sounds and applies the deterministic spoken commands: new line, new paragraph, and parentheses. What you lose is the language model's own work: lightly fixed punctuation, and filtered false starts and repeated words. You also lose the quote command, since only the model handles quotes.

How do you add structure by voice?

Say a command while you dictate and Parlo applies it as you speak:

1. Say "new line" to start a new line.
2. Say "new paragraph" to start a new paragraph.
3. Say "parentheses" and "end parentheses" around a side note.
4. Say "quote" and "end quote" around a quoted phrase.

New line, new paragraph, and parentheses run through a deterministic pass built into cleanup. Those three work even when the language model is unavailable, or when "Use AI cleanup" is off. Quote and end quote are different: only the language model converts them into quotation marks, so that command needs cleanup turned on.

How does dictation reach your feed from the keyboard?

Most dictations start from the keyboard, not the app. Tap the microphone key, and Parlo opens the app and starts recording. If you turned automatic start off in Settings, you tap record yourself once the app opens instead. A Live Activity shows on your Lock Screen and Dynamic Island while you record, and it ends once your text is ready.

When transcription and cleanup finish, you swipe back to the app you were dictating into. Apple requires that swipe: Parlo cannot return you there on its own. The keyboard watches for a signal the app sends the moment your text is ready, then inserts it into the field you left. A finished result also expires if you wait too long to swipe back, and the keyboard never inserts the same result twice.

What should you expect the first time you dictate?

Apple's on-device speech model sometimes needs a one-time download before it can transcribe. If that happens, Parlo tells you it needs the download and asks you to connect to the internet and try again. After that one download, transcription runs on the device from then on.

If you tap record and say nothing, or Parlo cannot make out any speech, you get an honest "No speech detected" message. Parlo never saves a blank recording to your feed as if it were a real dictation.

Does a custom dictionary fix a word dictation keeps mishearing?

No. A custom dictionary only governs the keyboard's autocorrect on typed text. Add a term in the iOS app, under Settings, Dictionary, Add word, and autocorrect stops changing it back when you type it. Editing the dictionary is an iOS feature; the Mac app has no dictionary. Dictation itself does not read the dictionary, so a name it mishears while you speak still comes out the same way next time.

Can you edit a note after it lands in the feed?

Yes. Tap any note to open it and edit it by hand, or run an on-device transform: Email, Bullets, Shorter, or Fix. Saving an edit keeps your originally dictated words stored for one-level revert. A transform you do not like is never the last word on your note.

What this means for a fast voice note

If you just want words on the page, cleanup does the small, mechanical work. No "um," no false starts, no doubled words, and lightly fixed punctuation. If you want more, spoken commands, your dictionary, and the on-device transforms take it further. All of it runs as on-device processing, with no account to create.