Voicy transcribes speech on the Mac it runs on. There is no account, no server, no analytics and no telemetry.
- Microphone audio is processed locally and is never sent to a transcription API.
- The last hundred transcripts, timing statistics, corrections and preferences
are stored in
~/Library/Preferences/com.helpsolution.voicy.plist; runtime state and unfinished recordings live under~/Library/Application Support/Voicy. - Saving dictated text is off by default. When "Save dictated text" is
switched on in Settings → Statistics, the text of each dictation is also
appended to
~/Library/Application Support/Voicy/stats/transcripts.jsonl, readable only by your user account. Nothing in that file expires on its own: switching the setting back off stops new writes and leaves what is already saved, which "Erase saved dictations" in the same panel deletes. - While saving is on, there are no per-app exceptions: anything dictated into a password manager or a banking app is saved like everything else. The bundle identifier of the app the text went to is recorded alongside it, regardless of the "Record applications" setting, which governs the statistics log only.
- Deleting an entry from the history overlay deletes it from that file too. It is deleted from Voicy — copies may still exist in Time Machine backups or APFS snapshots until those rotate, which no application can reach into.
- Diagnostic logs are written to
~/Library/Logs/Voicy.log, plus a launchd redirect at~/Library/Logs/Voicy-agent.launchd.log. Neither contains transcript text. - Audio for a recording in progress is written to disk only as a crash-recovery safeguard, and is deleted once the recording has been handled.
- The speech model is cached by FluidAudio under
~/Library/Application Support/FluidAudio/Models.
None of this is uploaded anywhere. Usage statistics are computed and displayed locally; removing the files above removes the data.
Voicy --mcp exposes the saved transcripts to an assistant over the Model
Context Protocol. It is read-only — it has no tool that writes, deletes or
changes a setting — and it speaks over stdio, so it opens no port and only
answers the client that launched it.
Voicy itself still makes no network request. But the client on the other end may be one that sends what it reads to a hosted model, and then the words leave the machine — at your request, in the same way that copying the file somewhere would. Two things follow, and both are deliberate:
- Nothing is exposed unless you switched saving on and connected the server.
- The search tool caps how much it returns and reports how much matched, and the statistics tool answers from counts alone, so a broad question does not drain the whole log into a conversation by accident.
Voicy opens exactly one kind of connection: the first-launch speech-model download that FluidAudio performs from Hugging Face. After that, dictation works with no network at all. There is no update check and no release feed — new versions are installed by hand.
- Microphone records speech while dictation is active.
- Accessibility finds the focused field and inserts the resulting text.
- Input Monitoring observes the configured global hotkey.
Voicy also uses AppleScript to mute the system output volume during recording, so audio from other apps is not captured back through the microphone. macOS may ask for automation permission for that; declining it only disables muting.