Offline dictation
on a Mac,
honestly.
What “offline” actually means
Dictation is offline when the speech model that turns sound into letters runs on your own machine. It is a real distinction, and it is worth understanding before you pick a tool — but it is not free. The models that fit comfortably on a laptop are not the models that get accents, noise and rare vocabulary right.
HearFlow makes that choice for you, and it is worth stating plainly. With a connection, dictation on a Mac is entirely cloud: the audio goes to api.hearflow.it, our own Cloudflare Worker, and from there to Groq (United States), where the largest Whisper model transcribes it. The text comes back and the audio is discarded — the transcription keeps no recording and no transcript (only Audios sync, if you turn it on, keeps dictations for 90 days). Without a connection, HearFlow falls back to the local model on your Mac, provided it has been downloaded: it is not part of the app download, so treat it as a safety net rather than a promise. On Windows the transcription always runs on your PC.
What this means for you
- Transcription keeps nothing. The audio is used to produce the transcript and then discarded. No recording or transcript is stored for transcription — only usage counters linked to your account, so your plan knows where it stands. Audios sync, if you turn it on, keeps dictations for 90 days.
- The route is written down. On Mac the audio goes to api.hearflow.it, our own Cloudflare Worker, and from there to Groq (United States) for transcription. That is the whole path, and it is in the Privacy Policy.
- Your text is yours. It is typed straight into the app you were already using. It reaches us only if you switch on sharing yourself — the improvement program or the MCP server.
- Your dictionary is local. The names and terms HearFlow learns live in a file in your user folder and are applied on your computer. They stay on your Mac unless you use the MCP server or a team dictionary.
- Windows is local. On a Windows PC the speech model runs on the machine, and nothing leaves it.
How it works on a Mac
A dictation app has three moving parts. A hotkey that captures the microphone — in HearFlow, ⌥Space, held for as long as you talk. A speech model from the Whisper family: on Mac, Whisper large-v3, reached over an encrypted connection; on Windows, a Whisper model running on the PC. And a text injector that types the result into whatever app has focus, which is why it works identically in Mail, Slack, Xcode and a browser text box.
Two things separate a good local setup from a merely working one. The first is a personal dictionary: the recogniser is told, before it starts, that “Beatrice” is a colleague and “pgvector” is a word — that single detail moves accuracy more than a bigger model does. The second is post-processing: punctuation, paragraphs and the removal of the false starts everyone makes when speaking.
When the local model steps in
If the transcription service cannot be reached — no connection, a limit, an error — and the local Whisper model has been downloaded on your Mac, HearFlow falls back to transcribing there. That model is not part of the download: it is fetched on first launch, so the fallback is a bonus rather than a guarantee. If you need dictation with the Wi-Fi off, check that the local model is installed before you rely on it.
What your Mac needs
- macOS 13 or later. Earlier versions lack the on-device audio APIs the app relies on.
- Apple silicon. M1 or newer. Windows builds are x64.
- 8 GB of RAM minimum, 16 GB if you want the local model always loaded.
- Around 500 MB to 1.5 GB of disk if you keep the local model for the fallback. Downloaded once, then never again.
- Microphone permission. The only permission the app asks for.
Three options,
honestly.
Not one of these is the right answer for everybody. Here is the short version, and where to read the long one.
Pick by the
job you have
Coming from Wispr Flow
Where the audio goes, what a subscription costs over three years, and the meeting notes job we do not do.
Read the comparisonComing from superwhisper
Local vs cloud by hardware, tone presets, the meeting recorder, and what one flat price buys instead.
Read the comparisonWhat HearFlow costs
Plans, what is included, and an ROI calculator that assumes nothing flattering.
See pricingSet up in three steps.
About four minutes, most of which is the model downloading while you make coffee.
Download and open
Drag HearFlow into Applications, launch it and sign in with email, Google or GitHub. macOS asks once for microphone access. The local speech model downloads in the background, for when the network is not there.
Teach it your words
Paste in the names, clients, products and acronyms you say every day. Twenty entries is enough to feel the difference on the first paragraph.
Hold ⌥Space and talk
Anywhere you can type. Release, and the text is there — punctuated, paragraphed, with your names spelled correctly, and no copy of it kept anywhere.
Technical questions.
Do I need Apple silicon?
Yes: HearFlow for Mac needs macOS 13 or later on Apple silicon, M1 or newer. Since transcription happens over the network, the accuracy is the same on an M1 Air as on an M4 Max.
On Windows, the build is x64 and the model runs on the PC, so there the hardware does affect speed.
How much RAM do I need?
8 GB is the practical floor and is fine with the compact model. 16 GB lets you keep the large model resident alongside a browser, an IDE and a video call without the machine swapping.
Which model does HearFlow use?
On Mac, Whisper large-v3, reached through our endpoint api.hearflow.it and run by Groq in the United States. On Windows, a Whisper model on the PC.
The local Mac model, used as a fallback when the service cannot be reached, is between 500 MB and 1.5 GB on disk and is downloaded on first launch.
Does dictating without a connection cost accuracy?
Some, yes, and that is exactly why the Mac app reaches for the big model first. A model small enough to sit on a laptop gives up ground on accents, background noise and rare vocabulary.
What moves accuracy most, either way, is context: a recogniser that knows your colleagues’ names beats a bigger one that does not, which is why the personal dictionary matters more than the megabytes.
Is my dictation stored anywhere?
No. On Mac the audio travels encrypted to api.hearflow.it and on to Groq for transcription; the transcription endpoint keeps neither the audio nor the text. All that is written down is a usage counter tied to your account; dictations are kept only if you turn on Audios sync (90 days, audio in the EU).
The text itself goes to your cursor, in the app you were already using.
Which tool should I pick?
If you need dictation across Mac, Windows, iOS and Android, or a meeting notetaker, Wispr Flow does that job. If the speech model has to stay on the machine no matter what, superwhisper does that job. If you write for hours and want the most accurate transcription with your own vocabulary and no recording kept by the transcription, that is what we built HearFlow for.
Try it and
read the route.
Fourteen days of Pro, no card. Then open the Privacy Policy and check that every hop we describe is the one you get.
Download for MacmacOS 13+ on Apple silicon · Windows x64 · from €12 / month