Dictate / Guides / On-device dictation
What on-device dictation means
On-device dictation means the speech model runs on your own computer, so your voice is turned into text without ever being uploaded. Here is what that changes, and how to check it in 30 seconds.
Two ways to turn speech into text
Cloud dictation. The app records you, uploads the audio, a server transcribes it and sends the text back. The vendor's privacy policy decides what happens to the recording. The app stops working when the connection does.
On-device dictation. The app ships (or downloads once) a speech model and runs it locally. On an Apple Silicon Mac that usually means the Neural Engine. The audio is processed in memory and can be discarded immediately. Nothing is uploaded, so there is nothing to retain, leak or subpoena.
What it changes in practice
- Privacy becomes structural. You don't have to trust a retention policy; there is no copy to retain.
- It works everywhere. On a plane, in a basement, on hotel Wi-Fi, behind a corporate firewall that blocks AI services.
- Latency is local. No round trip; the first word appears as fast as the chip can run the model.
- No account, usually. With no server to meter, there is nothing to log in to.
- The trade-off: the model has to fit on your machine. Current on-device models cover major languages very well and rare ones less so, and they don't rewrite your words the way cloud AI assistants can.
The 30-second test
- Turn Wi-Fi off.
- Dictate a sentence.
- If text appears, the app is on-device, at least for that language.
Watch for the half-way cases. Apple Dictation is on-device for about forty languages after a model download and cloud for the rest; Keyboard settings tells you which applies to you. Some third-party apps keep a local history but transcribe in the cloud; local storage is not local processing.
A worked example
Dictate, the app we make, is on-device by design. It downloads NVIDIA's Parakeet model (about 600 MB) on first launch and then transcribes 21 European languages on the Neural Engine, detecting which one you're speaking. Hold a key, speak, release: the text lands at your cursor, and the recording is gone. The privacy page lists every connection the app ever makes so you can verify it with a network monitor. Nothing you dictate goes to the cloud.
Other on-device options for Mac are compared in local speech to text on Mac.
Questions
What does on-device dictation mean?
On-device dictation means the speech recognition model runs on your own computer or phone, so your voice is turned into text without being sent to a server. Cloud dictation uploads the audio, transcribes it on the vendor's servers and sends the text back.
Is on-device the same as offline?
Nearly. On-device describes where the processing happens; offline describes whether a connection is needed. An on-device app works offline by construction, though most need internet once to download their model. The reverse isn't guaranteed: an app can keep your history locally and still transcribe in the cloud.
Is on-device dictation less accurate?
Not for everyday dictation in a major language on a recent Mac. Models such as NVIDIA Parakeet and OpenAI Whisper run on the Apple Neural Engine in a fraction of a second and are the same model families cloud services build on. Cloud still wins for rare languages and for AI rewriting of what you said.
How do I know if my dictation app is on-device?
Turn Wi-Fi off and dictate. If text appears, the transcription happened on your device. For Apple Dictation, Keyboard settings states whether your language is processed on the device. For Dictate, the privacy page lists every connection the app makes; none happen while you dictate.
Try on-device dictation
Apple Silicon Mac · macOS 14 Sonoma or later