What has to be on disk before you lose the network
Offline Mac dictation needs three things in place before Wi-Fi disappears: a downloaded speech model, microphone permission, and a working path for text to reach the field you are looking at. Miss any one and the app either refuses to start or quietly waits on a server. People remember the download and get caught by the insert path.
- A speech model on disk for every language you plan to speak.
- Microphone permission on the same macOS user account you will be logged into.
- Accessibility permission if the app inserts text by typing or pasting.
- Free disk for the model, which ranges from a few hundred megabytes to several gigabytes.
Auri Voice handles the first item with engines you download once: Fast Multilingual (Parakeet TDT v3), Fast English (Parakeet TDT v2), Low-Latency Multilingual (Nemotron 3.5 ASR), Lightweight Any Mac (Whisper Base), High-Quality Local (Whisper Large Turbo), plus Auri Cloud and built-in Apple on-device speech. Recognition runs on the Mac, and so does the cleanup pass that strips filler words, drops false starts, and adds punctuation. Cloud AI is a separate switch you can leave off.
The fifteen-minute airplane-mode test
Treat offline support as a claim you can check rather than a bullet point you trust. The test takes about fifteen minutes on trusted Wi-Fi plus five minutes with the radios off, and it tells you exactly which apps will cooperate at 30,000 feet.
- Download your engine and open the app once while you still have a connection.
- Grant Microphone and Accessibility, then quit and relaunch the app so the permissions take effect.
- Turn on airplane mode from the menu bar.
- Dictate two paragraphs into Notes, one reply in Mail, one message in Slack, and one paragraph in a browser document.
- Read the results. Look for missing punctuation, dropped last words, and text landing in the wrong window.
Write down which apps passed. If Notes works and Slack does not, that is a focus or permission problem in Slack, and reinstalling the speech engine will not fix it. Repeat the test after you enable any cloud feature, even briefly, because settings pages sometimes switch a network path back on in ways you only notice when the connection drops again.
Where Apple Dictation stops being enough
Apple Dictation sits on the keyboard, costs nothing, and can run on device for supported languages after speech assets install. For short bursts on a trusted machine, that is often all you need. The gaps show up when you dictate for a living rather than a sentence at a time.
| What you want offline | Built-in dictation |
|---|---|
| Long takes without cutting off | Tuned for short bursts |
| Custom vocabulary for names and products | No user-managed word list |
| Searchable transcript history | Not provided |
| One shortcut that behaves the same everywhere | Depends on keyboard and app |
| Cleanup of fillers and false starts | You edit by hand |
Language coverage matters more than feature lists when you travel. Confirm each language you speak has assets installed before you leave home Wi-Fi. Switching to a second language mid-flight without the pack turns a local workflow back into a failed network request, and there is no way to fix that from a seat.
Recognition and cleanup are two separate questions
Most disappointment with offline dictation comes from a split people do not expect. Audio to text is one step. Text to clean writing is a second step, and plenty of tools run the first locally while sending the second to a server. Offline, you get a raw transcript with every um, restart, and missing comma intact.
Ask any app two questions before a trip:
- Does recognition run on device after the model downloads?
- Does the polish pass that removes fillers and fixes punctuation also run on device?
Auri Voice answers yes to both by default. Local recognition feeds a local text improvement pass, so an offline draft still arrives with fillers removed, self-corrections resolved, punctuation in place, and lists formatted. Cloud models are available when you want them and are not required for cleanup. That single design choice is what makes a flight or a locked-down office feel like a normal writing session instead of a transcription cleanup chore you postpone.
Choosing an engine for a laptop you carry
Local recognition trades disk and memory for independence from the network, and the right trade depends on the machine in your bag. Smaller engines load faster, stay cooler, and keep pace with quick replies. Larger engines usually handle accents, names, and noisy rooms with fewer corrections.
- Thin laptop on battery: start with the smaller engine and judge it on correction count, not on a demo.
- Desk Mac with headroom: try Whisper Large for a week and see whether the fixes actually drop.
- Frequent names and product terms: Parakeet supports custom words in Auri Voice, which removes a whole category of repeat edits.
- Warm machine or spinning fans: switch down an engine size for short messages and save the large model for evenings on hotel power.
Measure with your own writing rather than a benchmark. Dictate the same three paragraphs, including two teammate names and one number, into each engine and count the edits you make afterward. A model that looked better in a quiet room often loses to a lighter one that matches your typing rhythm during a real workday.
Failures that look like the engine and are not
When offline dictation breaks, the speech model is rarely the culprit. Recognition can succeed while the destination app refuses to accept, save, or send the text. Knowing the difference saves you from redownloading gigabytes in an airport lounge.
| Destination | Offline recognition | What still needs a network |
|---|---|---|
| Notes or a local Markdown file | Works | iCloud sync of the file |
| Works | Sending the message | |
| Docs in a browser | Insert usually works | Saving and collaboration |
| Slack | Insert usually works | Delivery to the channel |
Keep one local capture file for travel days and paste upward once you reconnect. Put a timestamp in the title so you do not overwrite a newer paragraph with an older one. If insertion fails in a single app, click the field again and retry before touching settings. Electron apps such as Slack lose focus more often than native ones, and a second click fixes most of it.
FAQ
Does Apple Dictation work offline on Mac?
For some languages it can, once the speech assets are installed. Coverage varies by language and macOS version, so turn Wi-Fi off and dictate a paragraph before you count on it during travel.
Do local engines need the internet after the download?
No. Once the model files are on disk, recognition runs on the Mac. You need a network only for the first download and for any cloud feature you deliberately turn on.
How much disk should I budget?
Anywhere from a few hundred megabytes to several gigabytes per engine. Check free space before a trip, because a nearly full SSD is the most common reason a download fails the night before a flight.
Does cleanup still work offline, or do I get raw transcript?
That depends on the app. Some tools run recognition locally and send the text out for smart formatting. Auri Voice keeps the cleanup pass local by default, so filler removal and punctuation still work with the network off.
My mic works offline but no text appears. What broke?
Almost always focus or Accessibility, not the model. Click into the field before you speak, then check that the app still holds Accessibility permission after its last update.