How to dictate on a Mac without an account
Dictation does not need a login. It needs a speech model on your disk, a key to hold, and permission to type into other apps.
1. Pick the key
A modifier nobody else uses is the right trigger: Right Option, Right Command, or Fn. Holding a character key interferes with typing; a modifier does not. Decide between press-to-start / press-to-stop and hold-while-speaking — most people settle on press once to start and once to stop after a day.
2. Grant three permissions
Microphone (to hear you), Speech Recognition (to use the system model), and Accessibility (to type into other apps). All three are one-time, and all three live in System Settings → Privacy & Security. If an app asks for Full Disk Access or your password, that is a different kind of app.
3. Choose an engine, not a subscription
A local engine is either the one built into macOS or a model the app downloads once. Bigger models are more accurate and use more memory; a 0.6B-parameter model is a good default on any Apple Silicon Mac. The useful test is a paragraph in your own voice, not a benchmark table.
4. Give it your vocabulary
Every recogniser mis-hears names. A dictionary entry — the phrase it hears, the spelling you want — fixes a class of errors permanently, and on models that support it, the words are passed to the engine before it starts listening.
5. Decide where the text goes
Paste at the cursor is the behaviour people expect: one undo removes the dictation, and the clipboard comes back afterwards. Copy-only is the safer choice in apps that block synthetic paste.
Frequently asked
Can dictation work offline entirely? Yes, with a local model. Turn off Wi-Fi and try; that is the whole test.
Is Apple's dictation an account? No, but it is also not configurable, and its accuracy on names and jargon is why third-party apps exist.
What about Windows? The same setup exists there, with a different insertion mechanism. Inmost is macOS only for now.
Inmost does all five steps in its first run: one key, three permissions, a model measured on your voice, and no account. $39, once.