Dictation is not like recording a meeting. A meeting is one file, once. Dictation is every sentence you write, all day, in every app you type in — the email to your lawyer, the message to your doctor, the half-formed idea you would never say out loud to a stranger. Whatever the app does with that audio, it does it hundreds of times a day.

So “is this safe?” is a fair question to ask of a dictation tool, and the honest way to answer it is not with adjectives. It is with where the audio goes, said plainly, and a way for you to check. This guide covers what Wispr Flow does with your voice, what offline dictation should actually mean, which apps on Windows and Mac do it, and how to verify the claim yourself in about a minute.

Disclosure: I make Samurai Scribe, one of the apps below. I have tried to describe the others the way I would want mine described, and every claim about them links to their own pages, checked in September 2026. Prices and policies change — follow the links before you buy anything.

What Wispr Flow does with your voice

Wispr Flow is a good product, and it is popular for a reason: it is fast, it handles context well, and it works across Mac, Windows, iOS and Android. If you have seen people rave about dictating three times faster than they type, they are usually talking about Flow.

It also transcribes in the cloud. That is not an accusation, it is their own description. Their privacy page states: “Transcription always happens in the cloud to provide the best speed and accuracy.” The same page describes a Dictation Cloud Storage setting, which “controls whether Wispr stores Dictation transcripts, audio, and dictation history on its servers”, and a separate setting that lets you choose whether your “audio, transcripts, and edits” may be used to “evaluate, train, or improve AI models”.

Read those together and the shape is clear. You can turn off storage, and you can opt out of training, but the audio still leaves your computer to be transcribed. That is the part no setting changes, and it is the part that matters if your answer to “can this leave my machine?” is simply no — because of a client agreement, a hospital policy, an NDA, or your own preference.

For what it’s worth, Flow’s pricing is Free at 2,000 words a week on desktop, and Pro at $15 a month, or $12 a month billed annually, as of September 2026.

What “offline dictation” should actually mean

“Local”, “private” and “on-device” are marketing words, and several products use them while still sending audio somewhere. Here is the checklist I would apply to any of them, including mine.

It works with the network off. Not “works offline in a pinch” — pull your Wi-Fi, dictate a paragraph, and watch it appear. If it fails, it was never local.
It types into every app, not just its own window. A transcription app that makes you copy text out is a different product from a dictation tool that types where your cursor is.
The model lives on your disk. Speech recognition that runs locally means a downloaded model — a few hundred megabytes to a couple of gigabytes. If nothing was ever downloaded, the recognition is happening elsewhere.
No account, or at least no account required to dictate. A login is a sign that something server-side is involved in the path.
Your text is stored only where you can see it. History and word counts are useful; a copy of them on someone else’s server is a different thing.

The last one is worth being specific about. Local processing and local storage are separate promises, and a product can keep one while breaking the other.

Offline dictation apps for Windows and Mac

Checked September 2026, from each product’s own pages.

App	Price	Platforms	Where audio is processed	Types into any app	Meetings too
Samurai Scribe	Free · Pro $19 once (beta; $39 list)	Windows, Mac (Apple Silicon)	On your computer	Yes	Yes — speaker-labeled transcripts and summaries
Handy	Free, open source	Windows, Mac, Linux	On your computer	Yes	No
MacWhisper	Free · Pro €64 once	Mac only	On your Mac	Yes	Transcribes files; not a meeting recorder
Superwhisper	Free tier; paid plans, see their pricing	Mac, Windows, iOS, Android	On your computer, with caveats	Yes	Has meeting recording
Voice Access	Free, built into Windows	Windows 11 22H2+	On your PC	Yes	No
Wispr Flow	Free 2,000 words/week · Pro $15/mo ($12 annual)	Windows, Mac, iOS, Android	In the cloud, by their own description	Yes	Yes (Notetaker)

A few notes that the table flattens:

If you are on a Mac and you mostly transcribe files, MacWhisper is excellent and I would not try to talk you out of it. It does system-wide dictation as well, and €64 once is fair. It is Mac only.

If you want free and open source, Handy is the honest answer. No license, no tiers, and you can read the code. It does one job.

If you are on Windows and have not tried Voice Access, try it before you pay anyone. It is already installed, Microsoft says it “uses modern, on-device speech recognition to accurately recognize speech and works even without the internet”, and for straightforward dictation it is decent. It is a dictation and control tool, not a notes tool.

Superwhisper runs local models and supports Mac, Windows, iOS and Android, with one caveat worth knowing: their site notes that “Offline models only run really well on Apple Silicon macs”, and that “Intel Macs work best with Cloud models”. Check their pricing page for current figures — the plan structure changed during 2026 and I would rather send you to the source than print a number that is wrong by the time you read this.

Dictating privately on Windows and Mac, step by step

This part is specific to Samurai Scribe, since it is the one I can describe from the inside.

1. Install it

Download it for Windows 10/11 64-bit or an Apple Silicon Mac. No account, no sign-up. The Windows installer is code-signed, so Windows shows a verified publisher; on a new release SmartScreen may still ask you to confirm.

2. Pick a model

Setup asks what you want to use it for — meetings, dictation, or both — and only sets up what that needs. The speech model downloads once, to your disk. Everything after that runs on your own CPU or GPU.

3. Hold a key and talk

Hold fn on macOS, or Ctrl+Win on Windows, speak, and release. The words are typed into whatever app has focus. Double-tap the key instead and it keeps listening hands-free until you press again. A small on-screen pill shows when it is listening and working, and you can set it to always show, show only while dictating, or never.

You can choose between transcribing when you finish speaking, or live as you speak with a smaller model.

Dictation runs as a two-stage chain, and the app shows you which stage it is in:

Transcribe turns what you said into text; Polish tidies it up. Polish can be switched off.

4. Let it tidy up what you said

Speech has false starts in it. The Clean up polish prompt fixes those automatically and is free. Other polish prompts, including ones you write yourself, are Pro, and they can rewrite selected text anywhere — ⌥1–⌥5 on macOS, Win+Alt+1–5 on Windows.

5. Find your history

The Dictation page keeps what you have said, with word count and speaking speed. Like the rest of the app, it is on your machine.

How to check it is really offline

Here is the part I wish more people demanded, because it turns a marketing claim into something falsifiable.

Turn off your Wi-Fi. Dictate a paragraph. Watch it type.

That is the whole test, and it takes a minute. It works after the first run, once the model has downloaded. If a dictation tool passes that test, its recognition is genuinely happening on your computer; if it doesn’t, it isn’t, whatever the landing page says.

If you want to go further, run a network monitor — Little Snitch, or Wireshark — and watch what the app talks to while you dictate. Ours is documented in advance: every request the app makes, and what it carries, is listed on our network disclosure page. You should not have to take my word for any of this, which is the point.

One app for dictation and meetings

Most dictation tools do dictation. Most meeting tools do meetings. If you do both, you usually end up paying two companies and keeping your text in two places.

Samurai Scribe does both, with the same local models: hold a key to dictate into any app (see the interactive dictation demo), and record a meeting — online call or in-person conversation — to get a speaker-labeled transcript, an AI summary and plain Markdown files in a folder you choose. If the meetings side is what you care about, I wrote about that in detail in a Granola alternative for Windows that works offline.

FAQ

Does it really work with no internet at all? Yes, after the first run. The app downloads its speech model once, and dictation and transcription run on your CPU or GPU from then on. Left alone it still makes a daily check for updates and, if you bought Pro, a licence check — both listed on the network disclosure — and you can dictate with the network off entirely.

Does dictation work on Windows, or is this another Mac-only tool? Windows 10 and 11, 64-bit, and Apple Silicon Macs. There is no Intel Mac build yet.

Which languages does it handle? Whisper covers 99 languages. NVIDIA Parakeet, free too, has one model for English and one that detects 25 European languages automatically; Pro adds Parakeet Ultra, which misses fewer words in the same 25.

Is dictation free, or is it the paid part? Dictation is free, including the Clean up polish prompt. Pro ($19 once during the beta, $39 after) adds the three largest Whisper models, Parakeet Ultra, the other polish prompts and custom ones, automatic recording of calendar meetings, live meeting help and skills that run by themselves. To hear a Pro transcription model on your own audio before you buy, download it in Settings → AI Models → Transcription and try it in Compare models.

Will it type into any application? Yes — it types into whatever has focus, the same as any keyboard input, so your editor, browser, chat app and email client all work. It also has its own editor if you want one.

Is it as accurate as cloud dictation? Sometimes, not always, and it depends on the model you pick and the machine you run it on. Cloud services run larger models than most laptops can, and they have more context about what you are likely to say. The largest local models close a lot of that gap, and they cost you nothing per word. If raw accuracy matters more to you than privacy, a cloud tool may genuinely serve you better — that is a real trade-off and I would rather say so than pretend it away.