Offline dictation on a Mac
Almost every dictation app for the Mac describes itself as private, local or offline. Those three words are doing three different jobs, and the difference is the whole decision.
The three tiers
| Tier | What it means | How you can tell |
|---|---|---|
| 1 — Cloud | Your audio is uploaded and transcribed on someone’s server. Usually the most accurate in difficult conditions, because the model can be enormous. | It stops working with the network off. Their terms name sub-processors and a retention period. |
| 2 — On-device mode available | Transcription can run on your Mac, and often does. There is also a path to a server: for larger models, for cleanup, for licensing, for analytics. | It works offline but has cloud settings, an account, or a licence that activates. This is where most “local” apps actually are. |
| 3 — Never networked | There is no code path that sends anything. Not a setting you turn off — an absence. | A firewall log with nothing in it, over weeks. Or published source with no networking in it. |
Tier 2 is not dishonest and it is where most good software lives. It just cannot make the promise tier 3 makes, because a setting can be changed, a default can flip in an update, and a licence check is a network call however small it is.
Working out which tier an app is in
Marketing copy will not tell you, because all three tiers use the same adjectives. The method takes about ten minutes and is written out in full on the verification page — the short version:
- Compare the dates on the privacy policy and the terms of service. When they disagree, believe the terms.
- Run
nettop -p AppNameand dictate. Watch whether bytes move. - Install a firewall — LuLu is free — and read its log after a week of normal use. This is the step that finds analytics.
- Turn the network off and dictate. Note whether it degrades quietly rather than failing loudly.
Where the popular apps land
Checked on 21 August 2026, by reading the vendors’ own current documents and by measuring the shipped binaries. Verify before you rely on it — that is the point of the page.
| Tier | Apps |
|---|---|
| Cloud | Wispr Flow (their privacy page: “Transcription always happens in the cloud”), Willow by default, macOS built-in dictation for some languages and Macs. |
| On-device mode available | superwhisper, MacWhisper, VoiceInk, and most of the free open-source apps — several of which ship analytics enabled by default, with local transcription. |
| Never networked | A short list, and it is short because a licence check alone is enough to leave it. Vocula is built for this tier: no account, no licence server, no analytics, no update ping unless you ask for one. |
The question nobody asks, which matters more than the tier
Where does the transcript go after it is made?
An app can be perfectly offline and still keep every word you have ever dictated in a plain text file, unencrypted, indefinitely. Of nine local Mac dictation apps measured in August 2026, none encrypted its history — one stored transcripts as plain JSON inside its preferences file, another kept the raw audio of every dictation next to it.
Vocula keeps one encrypted file per day, sealed with a key held in your Keychain and tied to this Mac. The window is a year, history can be switched off entirely, and deleting a day deletes the file. That is a smaller promise than “offline”, and it is the one that protects you from the person who has your laptop.