What an AI dictation app actually does

An AI dictation app uses AI twice: once to turn your speech into text, and again to clean that text up. Those are two separate jobs, and vendors place each one in a different spot. Parlo runs both on your device, with Apple's speech engine for recognition and an on-device model for cleanup, free during the beta.

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

"AI dictation app" gets used for a wide range of products. Some just clean up punctuation. Others rewrite your voice note into a different kind of document. Apple's own dictation, built into iPhone and Mac, sits at the simple end. Parlo, Wispr Flow, superwhisper, Spokenly, and Monologue by Every each add something on top of basic speech recognition. Each one places its AI work in a different spot too, on the device, in the cloud, or split between the two depending on a setting. This guide separates the two jobs an AI dictation app actually does. It shows where each product runs each job, then answers the questions that matter before you trust one with your voice.

What is an AI dictation app?

An AI dictation app does two distinct jobs, and it helps to name them separately. The first job is speech recognition. That step turns the sound of your voice into a raw transcript. The second job is text cleanup, sometimes stretched into full rewriting. That step takes the raw transcript and fixes it up, or turns it into something else entirely, like a formatted email. Both jobs can be called AI. Both can run through a language model. They are not the same task, though, and a vendor's marketing page rarely draws the line for you.

Speech recognition on its own is not new. What changed recently is the second job. A language model can now clean filler words and fix punctuation. It can also follow a spoken instruction like "make this shorter," instead of a raw transcript landing on the page exactly as spoken. That second job is what most people mean when they say a dictation app uses AI today. The first job, recognition, has quietly used machine learning for years.

Where does the AI actually run?

This is the question that matters for privacy, and vendors answer it differently for each job. Some run recognition and cleanup entirely on your device. Some send audio to a server for recognition, then run cleanup there too. Some split the difference, with local recognition and an optional cloud model, or local recognition paired with cloud cleanup. The table below compares five products people search for alongside Parlo, using each vendor's own site.

Wispr Flow states plainly on its privacy page that transcription runs in the cloud every time. superwhisper and Spokenly both let you choose. You can pick a local model that never leaves your machine, or a cloud model through your own API key or the vendor's backend. Spokenly names its offline path a Local Only Mode. Monologue defaults to a cloud path, with a stated zero data retention policy from its AI providers, and adds a local transcription option on Mac. Parlo does not offer a cloud option at all. Both AI jobs, recognition and cleanup, run on your device every time, on iPhone and on Mac.

Is an AI dictation app as accurate as cloud dictation?

It depends on the engine, not on where it runs by itself. Parlo transcribes with Apple's SpeechAnalyzer and SpeechTranscriber. That is the same engine family behind the newest version of Apple's own system dictation, not a smaller model chosen to save bandwidth. Cleanup then runs through an on-device language model. It removes filler sounds and false starts. It also understands a spoken command like "new paragraph," adding structure as you speak. A custom dictionary is a do-not-correct list for the Parlo keyboard's autocorrect, so a name you type stays as typed; it does not change what the speech engine hears. Parlo publishes no accuracy number, and this guide does not rank the apps in the table on accuracy either. Accuracy also moves with the device, the room you are in, and how you speak, so one headline number would tell you little about your own dictation. It is also English-first for now, with no independent, apples-to-apples benchmark suite for this category to cite.

Where does your voice go when you use an AI dictation app?

With Parlo, nowhere it did not already ask to go. Audio is captured, transcribed, and discarded. No recording is ever written to disk, whether or not cleanup runs afterward. The first time Apple's speech model needs to download, that download needs a connection. Beta builds separately fetch updated transform prompts once a day, a request that carries no user data. Your finished transcript is stored locally, encrypted at rest. History sync is on by default. An encrypted copy travels through your private iCloud, so neither Apple nor Parlo can read it, and you can turn it off in Settings at any time. Vendors that run cleanup or recognition in the cloud are, by definition, sending your audio or your transcript to a server first. That is not automatically unsafe, but it is a different claim than on-device processing. It is worth reading a vendor's own privacy page rather than trusting a one-line pitch.

Is there an AI dictation app that is free, and what do the others cost?

Nothing, if you are using Parlo during the beta. There is no account to create and no subscription anywhere in the app. Every other product in the table above sells a paid tier. Wispr Flow's Pro plan runs $15 a month, or $12 a month billed yearly. superwhisper charges $8.49 a month, or a $249.99 lifetime purchase. Spokenly's Pro tier is $8.33 a month billed yearly. Monologue Pro is $15 a month or $144 a year. Three of those free tiers cap you somewhere: Wispr Flow by words per week, Monologue by words and by recorded notes, superwhisper by which cloud models you can reach. Spokenly's free tier is local models only, with no usage limit stated. Parlo's beta carries no such cap while it stays free, though a beta is not the same promise as a permanent price.

What does AI cleanup actually remove?

Cleanup is the second AI job, and on Parlo it is deliberately narrow. It strips filler sounds like "um" and "uh," trims false starts where you restarted a sentence mid-thought, and collapses repeated words. It can lightly fix punctuation and capitalization as it goes. What it does not do is summarize your note, invent new content, or change what you meant to say. A length guard checks the cleaned text against what you actually spoke. If the result comes back implausibly longer, Parlo falls back to a simpler, deterministic pass instead. The full mechanics of that pass, word by word, are laid out in voice notes to clean text.

Spoken formatting commands sit alongside cleanup. Saying "new line," "new paragraph," or "parentheses" and "end parentheses" works even with cleanup switched off. Those run through a deterministic pass that does not need the language model at all. The "quote" and "end quote" command is different. Converting spoken words into quotation marks needs the language model itself, so that one command only works with cleanup turned on. There is no spoken command for bullet points. The full list of what each command does lives in spoken formatting commands.

"Use AI cleanup" also defaults differently by surface. It is on by default on iPhone and off by default on Parlo for Mac, and you can flip either one in Settings. With cleanup off on either surface, Parlo still strips filler sounds and still applies the deterministic commands. What you lose without cleanup is the language model's own pass: lightly fixed punctuation, filtered false starts, and the quote command.

What else can an AI dictation app do with a saved note?

Past the raw transcript, an on-device transform can turn a note into something more specific. Tap a saved note in Parlo and edit it by hand, or run one of four transforms: Email, Bullets, Shorter, or Fix. Each one runs the same on-device language model used for cleanup, applied to a task instead of a light edit. Email assembles the greeting, body, and closing in code rather than leaving that structure to a model guess, and an optional "Sign emails as" setting fills in your name. Bullets turns the same words into a Markdown list, one point per line, without dropping any of what you said. Shorter cuts the note to roughly half its length, keeping the key points and your own tone. Fix touches only spelling, grammar, and punctuation, leaving the wording and structure exactly as you dictated them. Saving a transform keeps your originally dictated words on hand, so a change you dislike is never the last word on the note. You can revert back to what you actually said, once.

AI dictation on iPhone: Apple's keyboard or a Parlo keyboard

There are two shapes it can take. Apple's own dictation ships on every iPhone as part of the system keyboard, free, with no separate app to install. Parlo is a dedicated keyboard you turn on in Settings. Tap the microphone, speak in the app that opens, and swipe back to the keyboard, which inserts the cleaned text into the field you left. iOS does not let any keyboard extension record audio directly, Parlo included, which is why the recording step always happens in the companion app rather than the keyboard itself.

AI dictation on Mac: Apple's shortcut or a hold-Option hotkey

Apple's built-in dictation works on Mac from a shortcut you set in Keyboard settings. Parlo for Mac runs from the menu bar with a hold-Option hotkey. Hold the key, speak, and release, and the cleaned text pastes at your cursor in whatever app was frontmost. A double-tap on Option locks it hands-free for longer dictation, with a second double-tap to stop.

Where Parlo fits

Parlo transcribes with Apple's SpeechAnalyzer and SpeechTranscriber, running on your device. It is the only recognition engine Parlo uses, so there is no cloud fallback and no option to bring your own API key for a different one.

Cleanup runs through an on-device language model, built on Apple's FoundationModels framework. It strips filler words, false starts, and repeated words. A length guard blocks any pass that comes back implausibly longer than what you actually said, falling back to a simpler version instead.

New line, new paragraph, and the parentheses commands work even with cleanup switched off, since they run through a deterministic pass that does not depend on the language model. The quote command is the exception: it needs cleanup turned on, since only the language model converts spoken words into quotation marks. There is no spoken bullet command.

Use AI cleanup defaults on for the iPhone app and off for Parlo for Mac. Either default still leaves filler-word stripping and the deterministic commands running, since turning cleanup off only removes the language model's extra pass, not the whole pipeline.

Once a dictation lands in your feed, you can open it and edit by hand, or run an on-device transform: Email, Bullets, Shorter, or Fix. Saving a transform keeps your original dictated words on hand for one-level revert, so you can always step back to what you actually said.

Beta builds check parlokey.com once a day for refreshed transform prompts. That request carries no user data and has nothing to do with your audio or your transcript. App Store builds never make it.

Your audio itself is never written anywhere. Parlo captures it, runs both AI steps against it, and discards it, whether or not cleanup changes the final text.

Both apps are free during the beta, and neither one carries any purchase or subscription code. iPhone installs through the public TestFlight link, and Mac installs through a notarized download, with no account required for either install.