How to Choose a Private Dictation App

Split the app into three pipelines

Dictation privacy discussions go in circles because people treat an app as one thing. It is three, and each can be local or remote independently. Separating them makes vague marketing claims answerable.

• Recognition: microphone audio becomes text. Local requires a downloaded model.
• Cleanup: text becomes clean writing, with fillers removed and punctuation added. This is a second model and often a second location.
• Storage: transcripts and history persist somewhere, with or without sync.

Write those three rows for each candidate and fill them with yes or no, not adjectives. You will find apps that are local for recognition and remote for cleanup, apps that are remote for both, and a smaller set that are local for both. Once the table exists, the choice is about what you are comfortable with instead of about who wrote the better landing page.

Questions to ask before you install

These questions can be answered from documentation and settings in about fifteen minutes. If an app cannot answer them, that is information too.

1. Does recognition run on device after a model download, and for which languages?
2. Does the polish or formatting step run on device, or does it call a server?
3. What is the default state of every network feature on a fresh install?
4. Where is history stored, does it sync, and how do I delete it?
5. How long is anything retained on the vendor side, and can humans review samples?
6. Does the app work at all with the network off?

Question three is the one people skip and regret. An app can support a fully local mode that nobody enables. Question six is a fast sanity check you can run on day one, and it separates apps built to work offline from apps that merely mention local processing in a feature list.

Defaults tell you more than capabilities

Capability lists describe what an app could do. Defaults describe what happens on the machine of someone who installed it and started working. For privacy, only the second matters, because settings pages are visited once and forgotten.

Two examples from real products:

• Wispr Flow and Monologue both turn speech into clean writing, and that polish step commonly involves a cloud model. Monologue has a local mode where output sits closer to raw transcription, with smart formatting expecting a connection.
• Auri Voice on Mac runs local recognition and a local text improvement pass by default, so filler removal, self-correction handling, punctuation, and formatting happen on device without a switch to flip.

Neither arrangement is dishonest. They are different defaults with different consequences. If your work involves client names, health details, or anything under a data policy, choose the default that matches your obligations rather than the option you intend to enable later.

Platform limits explain most cloud defaults

It helps to know why defaults differ across devices, because the pattern is technical rather than philosophical. A Mac has disk and memory to host a speech model and a cleanup model. An iOS keyboard extension does not have the same headroom.

Use that pattern to place your work sensibly. Confidential drafting belongs on the Mac with a local setup. Quick replies on a phone can reasonably use a cloud default, and you switch to local transcription for the occasional message that warrants it. Splitting by sensitivity is more sustainable than trying to make one device do everything.

A one-hour evaluation you can run today

Testing beats reading. This sequence takes about an hour per app and produces evidence you can point at later, including in a conversation with a security reviewer.

1. Install, complete setup, and note every network feature that is on by default.
2. Dictate a fixed paragraph online. Save the output.
3. Turn on airplane mode. Dictate the same paragraph. Save that output too.
4. Compare the two for output quality as well as success. A big gap points to remote cleanup.
5. Reconnect and run a network monitor while you dictate and while you press each AI button.
6. Find the history store on disk and confirm you can delete it.

Keep the two saved outputs and the app version in a note. Behavior changes between releases, and a dated record lets you notice a shift instead of assuming nothing moved. Repeat the airplane-mode step after any update that mentions new AI features.

Match the app to the work, not to a slogan

The last step is deciding what you need, which is easier once the tables are filled in. Most people have a mix of work with different sensitivity, and the honest answer is different tools for different piles.

• Regulated or confidential: local recognition and local cleanup, no sync, history cleared after projects.
• Everyday work writing: local by default is comfortable, and an occasional cloud rewrite is a decision you make consciously.
• Personal messages and notes: cloud defaults are usually fine, and convenience wins.
• Shared machines: storage and retention matter more than the recognition path.

Write your own pile list before you shop, then pick tools that fit the top pile rather than the largest one. Auri Voice on Mac and Auri Keyboard on iPhone are independent apps, so you can run a local-by-default setup where it matters and a convenient cloud default where it does not, without pretending one configuration serves both.

FAQ