The short version
Mach captures audio only for an active dictation, streams it for speech recognition, optionally refines the transcript, and returns the finished text to the Mac app. The app then pastes that result into the text field you had focused.
Mach’s backend analytics are designed not to store raw audio or transcript text by default. Operational records contain metadata such as timing, status, selected language, mode, and word count.
1. The session becomes ready
Pressing Fn asks the backend to start a dictation session. The Mac app waits for a ready response before it begins sending microphone audio. That handshake prevents the app from recording into a session that cannot accept transcription work.
2. Audio is streamed for transcription
While the shortcut is held, the client streams audio over an authenticated WebSocket connection. Speech recognition turns that audio into text. Releasing the shortcut ends the input and asks the session to produce a final result.
3. Polishing is optional
In Refined mode, the transcript is sent through a language-model cleanup step that can remove filler, repair punctuation, and make the result easier to send. A transcription-only path can return the recognized text without that additional rewrite.
4. The Mac app pastes the result
The finished text returns to the native app, which restores the intended destination and inserts the result. Cancelling a dictation produces no final text and should record no usage.
What Mach records
Mach keeps the account and usage metadata needed to run quotas, subscriptions, reliability monitoring, and support. Dictation-event records are metadata-only by design; they do not contain the words you dictated or the raw microphone audio.
Cloud processing still means audio and, in Refined mode, transcript text must be processed during the active request. For the current policy and your data-control rights, read the privacy page rather than relying on a summary.