Offline dictation on a Mac
Most dictation apps that describe themselves as private are not offline. The two words get used interchangeably in marketing and they mean completely different things.
The ten-second test
Turn Wi-Fi off. Try to dictate a sentence. That is the entire test, and no privacy policy survives it.
An app that keeps typing runs its speech model on your Mac. An app that stalls, spins, or silently drops the recording was sending your audio to a server, whatever the landing page said about encryption, retention or compliance. Those things are all real and all beside the point — they describe how carefully someone handles a copy of your voice, not whether a copy exists.
Why “we don’t store audio” is a weaker promise
Zero-retention policies are common now and mostly sincere. But a policy is a statement about what a company currently chooses to do with data it already received. It can change with a new privacy policy, an acquisition, a subpoena, or a breach — and none of those require the app on your Mac to update.
On-device processing is not a stronger promise. It is the absence of a promise, which is better: there is no transmission to make guarantees about.
What runs offline on a Mac today
These process speech on your machine and keep working with the network off:
- superwhisper — $8.49/mo, $102/yr, or $249.99 lifetime
- Apple Dictation — Free — built into macOS
- MacWhisper — Free tier; Pro €59 (~$69) one-time; Pro Max $149
- Dragon — No Mac product. Dragon Professional v16 is Windows-only, $699.99
- VoiceInk — $29 / $49 / $69 one-time (1, 2 or 3 Macs); free from source under GPL v3
- Voibe — $7.50/mo, $59/yr, or $149 lifetime
- Apple Dictation — free and built in, on-device on Apple Silicon. Genuinely offline; see where it falls short.
And these do not, whatever else they offer:
- Wispr Flow — $15/mo, or $12/mo billed annually
- Aqua Voice — $10/mo, or $8/mo billed annually
- Willow Voice — $15/mo, or $12/mo billed annually
- Otter.ai — $16.99/mo, or $8.33/mo billed annually
- FreeFlow by Zach Latta — Free and open source (you supply API keys)
One caveat on the first list: several are hybrid, meaning they have local models and cloud models, and which one you get depends on the mode you are in. That is a real feature, but it means “does this work offline” becomes “does this work offline right now, in this mode.” Worth knowing before you rely on it in a tunnel.
The one thing offline apps still need a network for
Downloading the model. Speech models are large — FreeFlow’s recommended Parakeet V3 is 484 MB, and Whisper Large is 3.1 GB. You download once, and after that the app never needs a connection again.
Worth doing before you get on the plane rather than after.