Privacy statement
Last updated 2026-10-14
Owscribe turns your speech into text on your Mac and your iPhone. Recognition, and on a Mac the local language model if you allow it, run entirely on your device: your voice, your words, and what's on your screen are never sent to us or to anyone else. There are no accounts, no in-app analytics, no crash reporting to us, and no ads.
The sections below describe Owscribe for Mac; Owscribe for iPhone, which does less, has its own section after them.
What Owscribe keeps
- Your voice, only while you dictate. A recording starts with the dictation shortcut and stops when you release it or, if you tapped it, when you press it again (Esc cancels it); it also stops after ten minutes, when your Mac sleeps or locks, or when the microphone changes.
- Dictations. Each dictation's recording and transcript are saved on your Mac (with the wording as recognized, when cleanup changed it, and which cleanup did), with when it happened, how long it was, the outcome (for example "Pasted"), the name of the app you dictated into, the language it was recognized in, the language found before recognition and where, the contexts that applied (what you were writing, the project, the chat) and how each was picked, and the settings they changed. They stay until you delete them, one at a time in Dictations or all at once with Settings → Storage → Clear History.
- Microphone tests and practice use temporary audio that is discarded, never saved to Dictations.
- Voice calibration, if you run it, keeps only a few numbers per microphone (background noise level and the range of your voice's loudness and pitch), never the audio.
- Files you transcribe. An audio or video file you open or drop on Owscribe is read once and left as it is; its sound is kept as a recording in Owscribe's own storage, with the file's name and its transcript, until you delete that recording in Recordings or clear everything with Settings → Storage → Clear History.
- A diagnostics log of how Owscribe performed: when it launched, and for each dictation its outcome, how long each step took, how many words were recognized, the memory it used and the Mac's thermal state. Never what you said or wrote. It is removed with Settings → Storage → Clear History, and it leaves your Mac only if you send it to us.
- Your settings: your shortcut, your languages (started, on a new install, from the languages set in macOS) and the one used last, your vocabulary, cleanup choice and instructions, your contexts (kinds of writing, projects, chats) and what each changes, when and in which app each project was last detected, and what you allow on Settings → Privacy & Resources.
All of this lives in Owscribe's own storage on your Mac, which other apps can't read: ~/Library/Containers/app.owlive.owscribe.Owscribe. Deleting that folder, with Owscribe quit, removes everything Owscribe has stored.
What you choose to allow
Everything that reads from other apps, keeps more than the above, or uses memory and storage beyond dictating is a separate choice, in Setup and in Settings → Privacy & Resources, where each says what it does and what turning it off changes. Dictation works with all of them off: it records, transcribes, and lets you copy the text.
Reading from other apps also needs the Accessibility permission, which you can grant or revoke in System Settings at any time. With it alone, Owscribe can paste into the field you were in: it checks that the same text field is still focused and editable, never a password field, then pastes. To paste, it places the text on the clipboard for a moment, marked so clipboard managers don't keep it, and puts your previous clipboard back if nothing else changed it meanwhile; a password manager's copy is never put back. It never presses Return.
- Read the window you dictate into (on unless you turn it off): at the start of each dictation, it reads the window's title, the document, folder, or web address it shows, and, for apps that show an open project, that project's name, to pick what you're writing, the project and the chat. Used for that dictation, in memory; nothing is kept unless you also allow Remember where you dictate. Off: contexts apply only when you choose them from the menu bar.
- Use text near the cursor (off by default for new installs): it reads up to 500 characters before the cursor, up to 200 selected, and up to 150 after (an app that can only give a field's whole text has it read whole when it's under 4,000 characters, and only those parts used), so names and spelling match what you're writing, and, where a chat's language is looked for, which language you're typing in. Just before pasting it also reads the few characters beside the cursor, so the pasted text gets the right space and capital. Never kept.
- Read a conversation's language (off unless you turn it on; needs Read the window you dictate into): when you dictate into Messages, Slack (in a browser or its app), Discord or WhatsApp in a browser, or the Discord or Signal app, it reads the last five messages on screen, only to tell which language the conversation is in. The messages are never kept or given to a model; only the language is recorded with the dictation. Telegram doesn't make its messages available, so there the chat's language comes from the one last recognized in it.
- Remember where you dictate (off unless you turn it on): each dictation keeps the app, the window title, and the document, folder, or web address without its query or fragment, where tokens and search terms appear; never the text in the field or the messages on screen. Owscribe uses it to add a chat when you dictate into one whose window names it, and a project when an app shows the same project, or its repository's page is open, twice within 30 days; each chat keeps the language last recognized in it, which you can see and forget on its page. Each dictation also keeps your corrections, for Owscribe to learn from later: when you change its kind of writing or transcribe it again in another language, when you choose a context in the menu bar where Owscribe would have detected another, when you keep the original over the cleaned text, and edits you made after pasting (below). All of this stays on your Mac until you delete its dictation, or until you turn this off, which deletes all of it at once, along with projects seen only once. Chats and projects it added, and chats' languages, stay until you remove them.
- Notice edits after pasting (off unless you turn it on; needs Remember where you dictate): for up to a minute after a paste, it rereads the field around the pasted text each time the field changes, and keeps with the dictation what it pasted and what you changed it to. Only the pasted words and your changes to them are kept; the text before and after them, and anything you type after them, is read to find them and then discarded. It stops when you leave the field or the app, the text is gone, or you dictate again. Turning it off deletes the edits kept. Never in password fields or terminal apps.
- Learn from your edits (on when you turn on Notice edits after pasting, with its own switch): after five minutes without a dictation, Owscribe reads the edits kept and, where you replaced what it pasted with a name or term, recognizes that dictation's recording and up to ten recent ones again, on your Mac, to check the term. A term that doesn't change the others is added to the project, chat or writing style the dictation was in, or to All Apps once you correct it in a second one, with what it was pasted as, when, and which dictation taught it. Owscribe also keeps which edits it has read and the terms you removed, so they aren't learned again. Turning it off stops learning; terms already learned stay until you remove them.
- Local language model (off unless you allow it): a 2.3 GB download that runs on your Mac. It cleans up transcripts when Local model cleanup is chosen (everywhere or for a kind of writing), lays out emails and notes where you turn that on, and picks what you're writing when you allow that. It's given only the transcript, or for picking, the app's name, window title and address; it keeps nothing between dictations. Turning it off offers to delete its files.
- Keep the language model loaded: it stays in memory between dictations, about 2.3 GB more, so cleanup starts at once. Off, it loads while you speak and is freed after 10 minutes without a dictation.
- Pick what you're writing (off unless you turn it on; needs the local language model): when the app alone doesn't tell what you're writing, the model is given the app's name and identifier, the window title, the address or folder, and the project the app shows, if any, and answers with one of your kinds of writing. It keeps nothing.
Owscribe never reads text from password fields, password managers (1Password, Bitwarden, KeePassXC, Strongbox, Enpass, Proton Pass, MacPass, Keychain Access, Passwords), or terminal apps (Terminal, iTerm, Ghostty, Warp, WezTerm, kitty, Alacritty, Hyper, Tabby); for terminals it reads only the window's title and folder. A terminal inside an editor, such as VS Code's, can't be told apart from the editor. Outside a dictation it reads nothing from other apps, with three exceptions: while it notices edits after a paste, as above; when you press Check Now on a context, it reads the last window you used, once, and, if you allow reading conversations, how many messages are on screen and their language; and when you switch to an app where it once couldn't find the field you were in, it asks that app to turn on its accessibility support, because some apps (Chrome and apps built like it) only make their text fields available once asked. That request reads no content, and it's never made of a password manager.
Choosing what you're writing or a project from the menu bar needs no permission and reads nothing.
Files you choose
When you add a folder to a project, Owscribe keeps only the folder's path, to recognize windows that show a document inside it. It doesn't read, index, or change the folder's contents.
When you save a transcript (Save As, in Dictations or Recordings) or export your History (Settings → Storage → Export All), Owscribe writes only the files you asked for, where you chose: a transcript as text, Markdown, subtitles or JSON, or one Markdown file per dictation and recording with a history.json beside them that holds the same details Dictations shows (never the audio). It keeps no access to that folder afterwards.
Shortcuts
Owscribe's actions in Shortcuts and Spotlight (Start Dictation, Stop Dictation, Cancel Dictation, Get Last Dictation, Find Dictations, Clean Up Text) give a dictation's text only to the shortcut you run, and only when it asks for it. Nothing runs on its own.
Owscribe for iPhone
Owscribe for iPhone records, recognizes what you said on your iPhone, and copies the text for you to paste. It reads nothing from other apps or the screen, and it never reads your clipboard.
- Your voice, only while you dictate. A recording starts when you tap Dictate in the app, or from Owscribe's controls in Control Center, on the Lock Screen or the Action Button, or with Siri ("Dictate with Owscribe"); a control can start one while your iPhone is locked. It stops when you tap Stop, after ten minutes, or when a call or Siri interrupts it; Cancel discards it. While it records, a Live Activity shows the language and how long it's been, and afterwards how many words were recognized, never the words themselves.
- Recognition on your iPhone. Owscribe opens for a moment to recognize the recording, which needs your iPhone unlocked; until then the recording waits, with only its length and language noted. Fillers such as "um" are removed by simple rules on your iPhone; there is no language model on iPhone.
- Dictations. Each one's recording and transcript are saved in Owscribe's storage on your iPhone (with the wording as recognized, when the filler removal changed it), with when it happened, how long it was, its language and its outcome. Owscribe for iPhone doesn't keep which app you pasted into or anything about where you were. They stay until you delete them, one at a time in Dictations or all at once with Settings → Storage → Delete All. A dictation's page plays its recording on your iPhone. Like other apps' data, they're included in your iPhone's backups (to iCloud or your computer) unless you turn Owscribe off there; the speech models aren't.
- The clipboard. When a dictation is recognized, its text is copied for you to paste in any app, and it's removed from the clipboard after five minutes. With Only on this iPhone off (the default), your copied text also reaches your other Apple devices through Universal Clipboard, Apple's Handoff feature, as anything you copy does; turn it on to keep it on your iPhone. Copy Again, and Copy on a dictation's page or in Dictations, copy it again the same way.
- Audio files you choose or share. Transcribe a File in Dictations reads the audio or video files you pick in Files, once, and so does sharing one to Owscribe from another app, such as a voice message from a messaging app (iOS hands Owscribe a copy, which it removes once read); Owscribe keeps its own copy of the audio, converted, as a recording in Dictations, and leaves your file as it was. It's recognized on your iPhone while Owscribe is open, and kept, deleted and backed up as dictations are.
- Sharing and exporting, when you ask. Share sends a transcript's text to the app you choose. Settings → Storage → Export All writes every transcript, as Markdown files and a
history.jsonwith what each dictation's entry holds (not the audio), to the place you choose in Files; Owscribe keeps no copy of the export. - Siri and Shortcuts. Owscribe's actions (Dictate, Dictate in a language you name, Stop, Cancel) start and stop a dictation; none of them gives your text to Siri or a shortcut.
- Your settings: your languages, the one chosen on the main screen (or Automatic) and the one last recognized, your vocabulary (given to the speech model while it recognizes, to spell your terms), whether copied text stays on this iPhone, and recordings waiting to be recognized. Whether a dictation is recording is shared with Owscribe's controls, so Control Center can show it.
- A diagnostics log on your iPhone of how Owscribe performed: for each dictation its language, how long each step took, how many words were recognized and the outcome, with the iPhone's memory and thermal state, and the crash and performance reports iOS gives the app. Never what you said or wrote. It's removed with Settings → Storage → Delete All, and it stays on your iPhone; Owscribe has no way to send it anywhere. (If you share analytics with app developers in iOS's settings, Apple may send us crash reports for Owscribe, under Apple's terms.)
- Permissions: the microphone, asked for when you set Owscribe up or first dictate. Notifications are used only to say that a model finished downloading in the background, delivered quietly without asking first; you can turn them off in Settings.
Owscribe for Mac and for iPhone share nothing with each other: there's no sync and no account. Text reaches your Mac from your iPhone only through Universal Clipboard, as above.
Network
Owscribe connects to the internet only to:
- Download the speech model (about 2.5 GB), once, when you set it up.
- Download the local language model (about 2.3 GB), once, only if you allow it.
- Download the Armenian adapter (about 0.1 GB), which teaches the speech model Armenian, once, only if you add Armenian to your languages and download it.
- Check for updates, in Owscribe downloaded from its website, which asks the same server whether a newer version exists; the request includes your Owscribe version. Owscribe from the App Store doesn't check: the App Store updates it.
- Download a newer version of a model you already have, only when you choose Update; until then the one you have keeps working, and once the newer one is in place the older one is deleted.
Owscribe from the Mac App Store and Owscribe for iPhone download their model files through Apple's Background Assets service. Each model revision has its own asset pack, downloaded only when you choose Download or Update. The files are checked for their pinned size and SHA-256 before use. Apple handles the transfer and any download-service data under its own privacy terms; Owscribe doesn't send these requests to its Cloudflare download server or collect download analytics.
Owscribe for iPhone downloads only the speech model (about 2.5 GB, the same as the Mac's) and, if you dictate in Armenian, the Armenian adapter (about 0.1 GB). Apple can continue a transfer in the background while your iPhone is locked. Open Owscribe again to finish checking and preparing the files if the app was closed. Start on Wi-Fi waits for an ordinary network before beginning; after that, the system manages network use. The App Store updates the app.
Owscribe for Mac downloaded from the website gets its models and update information from downloads.owscribe.com, served by Cloudflare. Cloudflare retains sampled HTTP request metadata, including the IP address, time, file requested, and approximate country, for up to 31 days. R2 data access logging is disabled, and we do not enable Logpush or export these request logs. We do not use this metadata for advertising or to track you across apps or websites. Your recordings, transcripts, settings, and anything read from other apps are never part of these requests.
Children
Owscribe isn't directed at children. Its recordings, transcripts, and settings stay on your Mac or iPhone; download-request metadata is handled as described under Network.
Changes and contact
If this statement changes, the new version will be published here with its date, and the release notes will say what changed.
Questions: hello@owscribe.com