What stays on your Mac when you dictate

Dictation is an easy place to be vague. “Private” can mean the vendor promises not to train on you, or it can mean the audio never left the computer. Murmur is the second kind of claim, and it is a narrow one: the speech model is local. You hold a hotkey, you speak, and the text is typed into the app that has focus.

This note is the longer version of that paragraph, because the short version gets repeated until it sounds like more than it is.

What “local speech model” covers

On-device speech recognition means the model that turns sound into text runs on the Mac. Murmur’s model is that kind of model. We do not send the utterance to a Seven Hills server to be transcribed. We also do not name a particular open-source architecture on this site, because the product description is “local speech model,” not a model card.

The text then goes where you were typing. Mail, a notes app, a code comment, a browser text field: if it can take keystrokes, it can take the result. That destination is your choice. Murmur does not become the editor. The use case Dictate emails on a Mac starts by putting the cursor in the message on purpose.

History and stats stay in the app. They are a record of your dictation on that Mac, not a public profile. We do not describe a sync feature, so do not assume one.

What it does not cover

The app you dictate into has its own behavior. If you dictate a paragraph into a webmail composer, that paragraph is in the composer. Sending the email is a network act you take, and it has nothing to do with where the speech model ran. If you dictate into a chat app, the chat app does whatever it does when you type.

A custom dictionary and snippets are yours. They exist so names and set phrases come out right. They are not a claim that the model knows your company. You teach the words you care about. The use case Custom dictionary and snippets is the practical version.

Holding a hotkey is the gesture. We do not publish a words-per-minute figure, and we do not compare error rates with another dictation app on this page. Proofread. A local model can still mishear a name you have not added.

How to judge another dictation app

The Wispr Flow and Superwhisper pages are written the same way. We say what Murmur does. We tell you which question to ask the other product — where speech is processed, whether text is typed into the focused app, whether you can keep a dictionary — and we do not invent their answers. If a comparison page is more confident than its sources, it is the wrong page.

For the cluster of steps, use How to dictate on a Mac. For the definition that should stay stable when marketing shifts, use local speech model.