What on-device dictation actually means

Last updated 11 September 2026

The distinction that actually matters

Every dictation app does the same three things: capture audio, convert it to text, and put the text somewhere. The only architectural question worth asking is where step two happens.

In cloud dictation, your microphone audio is encoded and sent to a server. A large model there transcribes it and sends text back. This is why cloud apps can be extremely accurate and why they are almost always subscriptions: someone is paying for that server every time you speak.

In on-device dictation, a much smaller model file sits on your disk. The audio goes into it and never travels. Accuracy used to be the trade-off. On Apple Silicon, with models such as Parakeet and Whisper, it largely is not any more — see Whisper vs Parakeet.

How to check any app's claim in two minutes

1. Install and complete first run with the internet on Almost every on-device app downloads its model the first time you launch it. That download is expected and is not the same as uploading your audio. Let it finish.
2. Turn off Wi-Fi and unplug Ethernet Aeroplane mode on a MacBook is the quickest way. Verify it: the menu bar Wi-Fi icon should show as off, not just disconnected.
3. Dictate a sentence into any text field Notes or TextEdit is fine. Speak for ten seconds and stop.
4. Watch what happens Text appearing means recognition ran locally. An error, a spinner, or silence means it did not — whatever the marketing page says.
5. For the thorough version, watch the traffic Little Snitch or macOS's own nettop will show you which processes open connections while you dictate. An on-device app should be quiet except for licence and update checks.

What "on-device" does not promise

This is where honest explainers stop and marketing pages carry on. On-device recognition is a real, checkable property. It is also narrower than most people assume.

• It does not mean the app never uses the network. Licence checks, update checks, crash reports and analytics are all normal and all separate questions. Ask about them separately.
• It does not mean nothing is stored. Most dictation apps keep a local history so you can recover a dictation that went wrong. That history is on your disk, which is better than a server, but it is not nothing — especially on a shared or unencrypted machine.
• It does not cover optional AI rewriting. Plenty of apps recognise locally and then send the resulting text to a cloud model to tidy it up. That is a second, separate upload, and it is the one people miss.
• It does not mean the vendor cannot change its mind. Architecture is a decision, not a guarantee. The protection is that you can re-run the test.

Where the audio actually goes, step by step

It helps to see the whole path once. On a Mac running Vaitly Voice, a single dictation looks like this:

1. You hold a shortcut. macOS hands the app a microphone stream, which required your explicit permission the first time.
2. Voice activity detection trims the silence at each end, so the model is not asked to transcribe your room tone. See what voice activity detection is.
3. The trimmed audio goes into the speech model on your disk. This is the recognition step, and it is entirely local.
4. Your own vocabulary corrections are applied — names, brands and jargon you have taught it.
5. The text is placed into whatever field your cursor is in, through the same system input layer a keyboard uses.
6. A copy is written to a local history, on your disk, for the retention period you choose.

At no point in that list does a packet leave the machine. The exceptions are the ones we name plainly: the first-run model download, and an optional cloud rewrite if you deliberately configure one.

Is on-device always the right choice?

If you are weighing the two, offline vs cloud dictation sets out the trade-offs with no thumb on the scale.