Three questions that decide the answer
Privacy claims on download pages compress a lot of detail into one adjective. Replace the adjective with three specific questions, and every dictation app becomes easy to compare.
- Audio: does microphone audio leave the Mac for recognition, or does a downloaded model handle it?
- Text: after the transcript exists, does it get sent anywhere for cleanup, rewriting, or formatting?
- Storage: where do transcripts live, for how long, and do they sync anywhere?
Notice that the second question has no bearing on the first. An app can honestly say recognition runs on device while routinely sending your finished sentences to a server for polish. That is not deception, and it does mean the answer to what leaves my machine is yes. Ask all three questions before you decide whether an app fits sensitive work, and get the answers from settings and a network test rather than from marketing copy.
The second hop most people miss
Modern voice typing does two jobs. First, audio becomes text. Second, that text becomes clean writing: fillers removed, restarts dropped, punctuation added, lists formatted. The second job is where the pipeline often crosses the network.
Tools such as Wispr Flow and Monologue built their reputation on turning speech into finished writing, and that polish commonly involves a cloud model. Monologue offers a local mode, and in local mode you often get closer to raw transcription, with smart formatting expecting a network connection. None of this is hidden, and it is easy to miss when you read a page that says on-device and stop there.
Auri Voice is built the other way around. Local recognition feeds a local text improvement pass, both on by default, so the polished sentence that lands in your document was produced on your Mac. Cloud models exist as an option when you want them. Practically, that means turning off the network changes your speed slightly and does not change the quality of what you get.
Where transcripts live after you finish speaking
Storage outlasts every session, which makes it the most durable privacy question. A dictation app that never touches the network can still leave months of transcripts in a searchable database on a laptop other people use.
| What to check | Why it matters |
|---|---|
| History location | Local files stay under your control; synced history does not |
| Retention | Old drafts accumulate customer names and numbers |
| Backups | Time Machine and cloud backups copy the history too |
| Shared accounts | Anyone on that macOS login may see your drafts |
| Export and delete | You need a way to clear it, not only to read it |
Auri Voice keeps history on your Mac, which puts it inside whatever disk encryption and backup policy you already run. Set a habit around it: clear history after client projects, keep FileVault on, and check the setting after major updates. History is useful for recovering a paragraph an app lost, and it deserves the same care as a folder of meeting notes.
How to verify a claim in one afternoon
You do not need to read source code to test a privacy claim. You need airplane mode, a network monitor, and about thirty minutes of deliberate poking at features one at a time.
- Dictate a paragraph with the network on and note the output quality.
- Turn on airplane mode and dictate the same paragraph. Compare output quality as well as success.
- If the offline version is noticeably rougher, cleanup was probably happening remotely.
- Reconnect and run a network monitor such as Little Snitch while you dictate.
- Toggle each AI feature on and off and watch which ones create outbound connections.
Record what you find in a note with the app version, since behavior changes across releases. Also read the privacy policy for the two things a test cannot show you: retention periods and whether human review of samples is possible. A test tells you what happens today on your machine, and the policy tells you what the vendor allows itself to do in general.
Habits for work that actually matters
Configuration protects the pipeline. Habits protect the content. Even a fully local setup records whatever you choose to say out loud in a room with other people in it.
- Never dictate passwords, one-time codes, or card numbers. Type them.
- Step somewhere private before you speak customer names, salaries, or medical details.
- Keep cloud AI features off for regulated material, and know which switch controls them.
- Clear history at the end of sensitive projects rather than at the end of the year.
- Read the draft before sending, since a fluent wrong number is easy to miss.
If your employer has rules about where data may be processed, a local-by-default setup is the easiest one to explain to a security reviewer. You can point to a downloaded model, a local cleanup pass, local storage, and a test you ran in airplane mode. That is a concrete answer, and it holds up better than a screenshot of a marketing page.
Comparing two apps without guessing
When you are choosing between candidates, put them through the same short evaluation instead of comparing feature lists. The rows below are the ones that separate similar-looking products.
| Row | How to answer it |
|---|---|
| Local recognition | Dictate in airplane mode after the model downloads |
| Local cleanup | Compare offline output quality to online output quality |
| Optional cloud | Find the switch and confirm dictation still works with it off |
| Storage | Locate the history file or database on disk |
| Retention and review | Read the policy, not the landing page |
Fill the same five rows for each app and the decision usually makes itself. If a vendor leaves the cleanup row unanswered in its documentation, treat that as unfinished homework rather than a pass. The question is reasonable, the answer is knowable, and any app you trust with client work should be able to state it plainly.
FAQ
Does Apple Dictation on Mac send audio to Apple?
It depends on language and configuration. Some languages can be processed on device once assets install, and other paths involve Apple servers. Check the setting for your language rather than assuming.
If recognition is local, is the app private?
Not automatically. The transcript may still be sent out for formatting, rewriting, or AI features. Ask about both pipelines separately.
Where does Auri Voice do its cleanup?
On your Mac by default. Filler removal, self-correction handling, punctuation, and formatting run locally, so the polished text does not require a server.
Can I verify these claims myself?
Yes. Turn on airplane mode and dictate, then repeat with each feature enabled. A network monitor such as Little Snitch shows outbound connections while you speak.
What about transcript history?
History is where privacy quietly leaks. Confirm it is stored locally, know whether it syncs, and set a retention habit if you share the Mac or back it up somewhere else.