The cloud path

Your Mac records audio, compresses it, and uploads it. A server transcribes it, often runs a language model over the result to punctuate and clean, and sends text back. The app inserts it. Along the way, your audio and your text exist on machines you do not control, for as long as that provider's policy says, and the app needs an account to know whose audio it is.

This path has real advantages: the provider can run models far larger than a laptop can, it can update them without shipping an app update, and it can support a hundred languages at once.

The on-device path

Your Mac records audio, transcribes it with a model that lives on the Mac, cleans it up with rules or a local model, and inserts it. Nothing is uploaded. There is no account, because there is nothing to attribute.

Its constraints are equally real: the models are the ones that fit on your machine, the quality depends on your hardware, and platform support follows the platform's engine. The two local engines people ask about.

What changes for you

  • Privacy: on device, a sentence you dictate is never anywhere but your Mac. This is architecture, not policy.
  • Latency: a round trip to a server has a floor set by the network. On device, the floor is the model. DeftNib measures 88 ms at the 95th percentile for short dictations, from key release to text.
  • Offline: on device works with Wi-Fi off. Cloud does not work at all.
  • Accounts: on device needs none. Cloud needs one.
  • Cost model: cloud costs the provider money per minute, which is why cloud dictation is a subscription. On device costs the provider nothing per use, which is why it can be sold once.

How to check what an app does

You do not need to trust a page like this one. On a Mac: turn Wi-Fi off and dictate. If it works, transcription is local. Then look at the app's own network list, if it publishes one, and compare it with what a firewall like Little Snitch or LuLu shows. An app that enumerates its calls (activation, updates, a counter) is telling you something an app that says "we take privacy seriously" is not.

Where DeftNib sits

Fully on the on-device path: Apple's speech engine and Apple's local model, on your Mac. The app makes four kinds of calls to our servers, listed with what each carries on the private dictation page, and none carries your voice or your text. Your Mac also downloads Apple's speech models and, when you accept an update, the installer. Those come in; nothing of yours goes out.

Mac with Apple Silicon and macOS 26 or later. Free up to 4,000 words a week, no account.

By downloading you agree to the Terms and the Privacy Policy.

Questions

Is on-device dictation less accurate?

It depends on the engine. Apple's SpeechAnalyzer on macOS 26 is the same engine behind Notes and Voice Memos; DeftNib adds rule engines and a local model on top. For English and Spanish the cleanup is the differentiator, not the raw transcription.

Does "on device" mean no network at all?

No app is honest saying that. Updates and licenses use the network. What matters is that voice and text do not.

What does DeftNib send?

Four kinds of calls to our servers, none carrying your voice or your text. The private dictation page lists each one with what it carries.