What leaves your Mac

This page describes normal processing and component downloads. The terms for voluntarily sharing data are in the privacy policy.

Processing on your Mac

Recognition, translation and text processing run on your Mac, without sending them to the cloud. Transcripts are kept in local history. You choose whether meeting audio is saved to disk.

Smart Keyboard (a paid feature) sees the keys you press so it can fix a word typed in the wrong layout or with a typo — only with the Accessibility permission. Typed text is not recorded in history or logs and never leaves your Mac; words you undo are kept only in the local exception list; the details are in Help.

Normal operation and downloads

Account operations, component downloads and statistics are not used for cloud processing of your speech.

Your account. Sign-in, registration, token requests, and changes to your name or password when you make them go to our server. Registration sends your email, password and an install-key signature.

When you are signed in, the sign-in, each licence-token refresh (once a day, or after four hours of use) and the start of a trial also carry your macOS language, the app version and how long the app was in use since its previous refresh; the trial start also carries a one-way fingerprint of the Mac’s hardware (a hash of its hardware ID; the ID itself is not sent). The server records the IP address, country and city the request came from. This keeps it to one trial per Mac and at most five devices per account, and shows when one account is used by several people. IP address and city are kept for 90 days, the rest for as long as the account exists.

Usage counters — only with consent. If you allow statistics, the app sends numbers, never text. What is in them: how many dictations and in which language or language pair; how long you spoke; how many characters came out — the count, never the characters; how long recognition, translation and text processing each took; which paid feature was blocked; and your country, taken from the system region. Two doors: by session when you are signed in, and when you are not, by a stable device id plus a signature made with the install key.

If you did not allow statistics, none of this is sent at all.

Update check — once a day, and you can turn it off. The app reads the version list cdn.verbaide.com/downloads/appcast.xml and compares a number. Nothing about you or what you dictate is sent. A new version is installed only after you press Install.

Translation dictionaries from cdn.verbaide.com. The catalogue is fetched when Settings are open, or when the language screen of the first-run setup is open. The files themselves are fetched when you download them. Both requests carry no cookies and no authorization header. Only the ru/uk→en dictionary ships inside the app; every other direction, including en→ru/uk, is downloaded.

Speech-model registry and models. The recognition/parakeet.json pointer and its manifest on cdn.verbaide.com are checked at launch, at most once a day. Offline, a cached or bundled manifest is used. The main model is downloaded from the CDN, with huggingface.co as the fallback source. The CTC term model (98 MB) downloads from Hugging Face in the background on first dictation. An installed model is updated with your agreement.

Speaker-labelling model. Downloaded from huggingface.co only when you press Download. Speaker labelling is experimental.

Local language-model metadata from cdn.verbaide.com/intelligence. Catalogue when Settings are open, manifest when a download starts. Metadata only — the weights are not on that host.

Local language-model weights from huggingface.co/Qwen/…. Downloaded by the Mac of whoever chose to install the local model, the same way the speech model is.

Term judge model — cdn.verbaide.com/intelligence/term-judge/… (1.8 GB), only on a paid plan or the trial; it downloads by itself in the background. There are two triggers: the start of a dictation, when dictation improvement is on and your lists hold at least one term; and Smart Keyboard, when it starts listening to the keyboard (the feature is on and Accessibility is granted). The model is not downloaded at app launch, and never on the free tier. Both triggers use the same address.

What the free tier does

Dictation and term dictionaries are free forever, with no word cap and no card. Download the speech models first; afterwards dictation works offline. Component catalogues fetched while Settings are open and the speech-registry check at launch do not depend on your plan.

Where to check this yourself

Point a proxy at the app and watch. Every address above is in CLAUDE.md in the source tree, and the network policy that decides whether a request may leave at all lives in the Rust core, in one list shared by macOS and Windows.