Offline dictation app: what offline really means

An offline dictation app should turn speech into text without sending your audio to a server, but few are offline for every step. Recognition, cleanup, and history can each run on the device, yet a first run still needs a one-time model download. Parlo works this way: on-device by default, honest about that one download.

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

People search for an offline dictation app for one of two reasons. One is a flight or a job site with no signal. The other is a wish to keep their voice off someone else's server. Those two reasons point at different features. Working without a signal is about the recognition step alone. Keeping your voice private is about every step: recognition, cleanup, storage, and sync, each of which a vendor can run locally or not, on its own. A page that says only "offline" without naming which step it means is not telling you enough to judge it.

A useful habit when reading any product page in this category is to ask, for each step, whether the page names an engine and a location. Some pages give you only a vibe. "Works offline" without saying whether that covers cleanup, history, and settings sync is a marketing summary, not a technical claim you can check. A page that names the exact engine and the exact step gives you something you can verify. What still needs a connection matters too, on a plane or anywhere else.

The players in this category split roughly into three groups. System dictation, built into iPhone and Mac, switches between an on-device path and a path that sends audio to a server. Which path it uses depends on the device, the language, and a setting. Hotkey and keyboard apps like superwhisper, Spokenly, Monologue, and Wispr Flow each pick a different default. Some process locally by default, some send audio to a server by default, and some let you choose per dictation. Parlo is the fourth shape: one on-device path, with no cloud model to switch to.

What an offline dictation app can honestly promise

Recognition on the device is the easiest promise to keep and the easiest to verify. A vendor either names the engine it uses and states plainly that it runs locally, or it does not. Parlo transcribes with Apple's SpeechAnalyzer and SpeechTranscriber, the same on-device engine family behind the newest system dictation, and it is the only engine Parlo uses. There is no cloud fallback and no bring-your-own-key option to a different service.

Cleanup on the device is a second, separate promise. Turning speech into readable text often needs a pass that removes filler sounds and fixes false starts. It also applies a command like "new paragraph." That pass can run on a server even when recognition ran locally. Doing so quietly breaks an offline claim without breaking the recognition claim. Parlo's cleanup runs on-device too, with an on-device language model, and a length guard that rejects a rewrite that looks implausibly longer than what you actually said.

No account and no server copy of your content round out what offline can mean beyond processing. An app that requires sign-in before you can dictate is keeping a record of you somewhere, even if the audio itself stays local. Parlo has no account and no server holding your content. Sync is keyed to your own iCloud identity, and signing out leaves a fully local mode with nothing to turn off.

What offline cannot honestly mean

The one exception almost every on-device app shares is the first run. Apple's speech model, and most other local models, need a one-time download before they can recognize anything. That means the very first dictation on a new device usually needs a connection. After that download completes, transcription runs locally without repeating it. A vendor that claims to work with no connection at any point, ever, including the first minute after install, is describing something none of these products actually do. How on-device recognition works, engine by engine, is its own explainer at on-device speech recognition.

A second honest exception is anything an app fetches that is not your audio or transcript. Parlo's beta builds refresh their text-transform prompt text once a day from parlokey.com. That fetch is a small one, carries no user data, and has nothing to do with a specific dictation. App Store builds skip this fetch entirely. It is worth naming because a strict reading of "no connection, ever" would flag it, even though nothing about your voice or your words is involved.

Sync is the third case, and it starts once you sign in with Apple. History sync moves your encrypted transcript to your private iCloud so it shows up on your other devices, and that step obviously needs a connection when it runs. It is on by default in Parlo, and you can turn it off in Settings, or skip signing in with Apple entirely to stay fully local. None of this changes where the speech-to-text step itself happens; it only affects what happens to the text afterward, if you choose to sync it at all.

How offline dictation apps compare

The table below pulls each product's own stated processing location, free tier, supported platforms, and price from its own site, checked the same day this guide was written. Read the "Where processing happens" column carefully rather than the product name alone. Two apps that both mention offline use can still differ on whether that applies to recognition only, to cleanup too, or to the app's every feature.

How Spokenly and superwhisper describe their own offline modes

Two of these products name offline processing directly on their own pages, and reading their exact wording is useful for judging any offline claim, including Parlo's. superwhisper's home page states plainly that it "works offline, so you can transcribe anytime," with a line about needing no Wi-Fi. Spokenly names a specific "Local Only Mode" that it says can restrict network access for full offline use. That mode layers on top of its separate local-model option, which already runs without sending audio to its servers. Parlo compared with Spokenly sets the two side by side.

Both descriptions hold up as far as they go. Local recognition models genuinely do not need a live connection to convert speech into text once the model is present on the device. Both claims describe the state after a local model is already on the device. Getting it there is a network event, whatever the page says about what happens afterward. Neither page says whether the claim covers every other feature, such as syncing settings or checking for updates. A claim about one step, stated clearly, is more useful than a broad claim about the whole app. This guide tries to hold Parlo to the same standard: name the step, then say what is true about it.

A claim of no connection at any point, ever, is one worth double-checking on any product's page, this one included. Every local model has to arrive on the device somehow, and that arrival is a network event even when everything after it is not.

Where does your audio go if you dictate offline?

Nowhere. Parlo captures audio, turns it into text on your device, and discards the audio once transcription finishes. No recording is written to disk and none is sent anywhere, offline or online. One thing can still leave your device afterward: an encrypted copy of your history, since sync is on by default once you sign in. Editing a note, or running an on-device transform on it, stays on the device like the transcription did.

Your local history itself is encrypted at rest under a key that stays on the device. Deleting a note moves it to a local trash for a period rather than erasing it outright. That trash is never synced, so it stays on the one device where you deleted it. There is no separate audio archive anywhere to also go delete.

Is on-device dictation accurate enough?

Accuracy is a separate question from where processing happens, and it is worth treating that way. Parlo transcribes with Apple's on-device speech engine, the same engine family behind the newest version of system dictation on iPhone and Mac. It then cleans the result with an on-device language model that strips filler sounds and false starts. Spoken commands like "new paragraph" work with cleanup on or off. Parlo publishes no accuracy percentage, and this guide will not invent one for anybody else, so a specific number is not something an honest comparison can offer. Which local engines actually run on a Mac, and how they differ, is covered in Whisper vs Apple speech.

What is worth checking instead is whether an app names its actual engine and lets you try it against your own voice and vocabulary. superwhisper, Spokenly, Monologue, and Wispr Flow each offer a custom dictionary; Parlo's is narrower, a do-not-correct list the Parlo keyboard's autocorrect reads, which does not change recognition. Parlo dictates in English for now, and transcription falls back to en-US, so tell the team about anything that trips it up.

Whether an app runs on-device or in the cloud does not by itself decide how accurate it feels to you. A cloud model can be trained on more data than fits on a phone. An on-device model skips the round trip instead, which can cut the delay before text appears. Neither fact tells you how well either one handles your accent, your industry terms, or a noisy room, which is a question only your own trial run answers.

Is a beta offline dictation app safe to rely on?

It is fair to ask, since Parlo is a beta. Dictation, cleanup, and the on-device pipeline work every day for current testers, and any known rough spots are listed on the beta page rather than hidden. Your history stays encrypted on the device no matter what a specific build does elsewhere. If sync is on, the copy in your private iCloud is encrypted the same way. A bad build cannot expose notes that were never stored in a readable form to begin with.

Parlo for Mac checks once a day for a new build and tells you when one is out, so staying current does not require watching a release page yourself. Feedback goes to beta@parlokey.com, and it shapes the next build, including anything about offline behavior specifically that does not match what this guide describes.

Where Parlo fits

Parlo transcribes with Apple's SpeechAnalyzer and SpeechTranscriber, and that is the only recognition engine it uses, with no cloud model waiting behind a setting.

Cleanup also runs on the device, using an on-device language model that removes filler words and false starts. It checks its own output against a length guard, so an implausibly long rewrite falls back to something simpler instead of being trusted blindly.

Audio itself is never written anywhere. It gets captured, turned into text, and then discarded, on every dictation, offline framing or not.

History sync, on by default once you sign in, moves through your own private iCloud end-to-end encrypted. Neither Apple nor Parlo can read the plaintext or hold the key. Turning it off, or simply not signing in with Apple, keeps everything local instead.

There is no Parlo account and no server of Parlo's own holding a copy of your content. Whatever sync you do use is tied to your personal iCloud identity, and signing out drops you into a fully local mode with nothing left to disable.

The one connection Parlo's on-device path still needs is the same one every local model needs: a one-time download of Apple's speech model on first run. After that, everyday dictation does not repeat the request.

Beta builds also make one small daily fetch of text-transform prompt text from parlokey.com, which carries no user data and has nothing to do with any specific dictation. App Store builds do not make this fetch at all.

None of this costs anything to try. Parlo's beta is free on iPhone through TestFlight and on Mac through a notarized download. Neither app carries a purchase code or a subscription tier waiting behind a trial.