Your voice and dictation
Talk uses microphone audio while you choose to dictate. Audio is held in memory and processed by the speech model on your Mac. Talk does not upload your audio or transcripts to Scribed for speech recognition, and it does not save audio recordings to history.
When you insert, copy, or export a transcript, the destination app, clipboard, or exported file receives that text. The destination app may store or sync it according to its own settings.
To paste a dictation, Talk temporarily saves your existing clipboard contents in memory and restores them after insertion, unless you have copied something new. Live typing inserts text with keyboard events instead.
History and preferences stay on this Mac
Optional history contains the transcript, its date and duration, and the original model text when cleanup changes it. History, vocabulary, snippets, and preferences are kept locally and separated by Scribed account. Talk does not sync them to your Scribed workspace.
Local history is not encrypted by Talk. You can choose how many entries to keep, delete individual entries, clear all history, or turn off new history in the app. Turning off history or signing out does not delete previously saved entries. Copies you export or put in other apps remain there until you remove them.
Your shared Scribed account
Talk requires a Scribed account. Sign-in and session checks contact Scribed over an encrypted connection. The app stores its session token in macOS Keychain; it does not persist your password or temporary two-factor verification cookies.
An internet connection is required to verify your session each time Talk opens. Once verified, dictation can continue offline for up to 24 hours in that running session. Signing out clears the app’s active account session and cancels pending dictation.
Shared-account information and requests about that information are covered by the Scribed Privacy Policy. This page explains the Mac app’s local processing; it does not mean the shared-account service operates offline.
Permissions and the model download
Microphone permission lets Talk hear your dictation. Accessibility permission lets the dictation shortcut work and lets Talk insert text into other apps. It checks the focused app, text field, and selection position to avoid inserting into the wrong destination; it does not read your document contents. You can review or revoke either permission in macOS System Settings → Privacy & Security.
The speech model downloads separately, about 500 MB, and is cached on your Mac. That download contacts the model-hosting service; your dictation audio is not part of the download request.
Diagnostics and support
Talk keeps a local support snapshot with permission and shortcut readiness, recording sample counts, and text-delivery status. This snapshot contains no audio, transcript text, document contents, or general keystrokes. It is not automatically uploaded to Scribed.
If you contact support, Scribed receives the message and files you choose to send. Do not include passwords, verification codes, or dictation content you do not want to share. For help or a privacy request, email hello@scribed.ai.