Private dictation on a Mac: what it means and how to check

Every voice app claims to respect your privacy. The claims mean very different things - and a few of them can actually be verified.

Written by the Bellow team

Three claims that sound identical and aren't

Read voice-app marketing and you'll meet the same phrases. They describe fundamentally different architectures.

"We don't store your recordings"

Your audio still travels to their servers and is transcribed there. The promise is about what happens after. It may be entirely sincere - but it's a policy, and policies depend on the company keeping them, its subprocessors keeping them, and nothing going wrong in transit or storage.

"Your data is stored locally"

This one is slippery. It frequently refers to settings and history being kept on your machine while the transcription still happens in the cloud. Local storage of results says nothing about whether your voice was uploaded.

"Processing happens on your device"

This is the strong version: the speech model runs on your hardware and the audio never leaves. It's the only one of the three where a network failure can't leak anything, because there's no transfer at all.

Why it matters more than it seems

Dictation is unusual: you use it for the things you'd type, which means it sees your most sensitive text. Client notes, medical context, salary conversations, unreleased work, half-formed opinions about colleagues. It's an unusually complete picture of your working life.

For some people it's also a compliance question. If you handle regulated data, "the audio was transmitted to a third party but they say they deleted it" is a much weaker position than "the audio never left the device."

How to evaluate an app in five minutes

1. Airplane-mode test. As above. Fastest signal available.
2. Look for an account requirement. Mandatory sign-in usually implies server-side processing.
3. Read for "zero retention" vs "on-device." The first admits transmission; the second denies it.
4. Check the install size. A local speech model is large - often 1 GB or more. A very small installer that transcribes well is almost certainly calling an API.
5. Ask whether it's checkable. Can the app show you evidence, or does it only assert?

Making a claim you can inspect

This is why Bellow ships a Privacy Audit. The app calls into macOS to read its own code-signing entitlements at runtime and displays them - including network access, none.

The reason that's meaningful: entitlements are resolved by the operating system from the signed binary on disk. It isn't a value we store in a config file and print back to you. Changing what it reports would mean re-signing the application. Combined with the fact that Bellow works with no internet at all, the claim stops being something you have to take on faith.

What Bellow does with your text

• Audio is transcribed on the Neural Engine and discarded. It's never written to disk.
• Transcripts are stored locally and encrypted with AES-GCM. The key lives in your Keychain, marked device-only, and never syncs to iCloud.
• Confidential mode stops history being written at all - dictate, paste, nothing persists.
• Export gives you everything as Markdown or JSON whenever you want it.
• No accounts, no telemetry. There's no analytics pipeline, so there's nothing to opt out of.