Release notes

0.12.0 · 2026-10-14

  • Owscribe for iPhone is coming soon to the App Store. Dictate in the app, or start from Control Center, the Lock Screen, the Action Button or Siri without opening it; stop, and the text is recognized on your iPhone and copied, ready to paste. It uses the Mac's speech model and your languages: Automatic recognizes which you speak, or choose one, and Dictate in… keeps a control for one language. Every dictation is kept in Dictations, to listen to, correct, retry in another language, share or export. Audio files, and voice messages shared from apps such as Telegram or WhatsApp, are transcribed too.
  • Numbers come out as digits: “48 hours” and “Windows 10” where the speech model heard a number but wrote it in words, in English, Armenian, Russian, German, French, Spanish and the other languages written with spaces between words. Words like “one of those” or “un café” stay as they are. Turn off Numbers as digits in Settings → Recording on the Mac, or under Transcription on the iPhone, to keep numbers as the model writes them.
  • Owscribe's home is now owscribe.com, and its address hello@owscribe.com. Updates and models come from downloads.owscribe.com; the old addresses forward there.
  • Owscribe learns names and terms from your edits. When you replace a word it got wrong after pasting, such as “FF impact” with “ffmpeg”, it adds the term to the project, chat or writing style you were dictating in, and to All Apps once you correct it in a second one. It works after five minutes without a dictation, checks each term by recognizing the recording and a few recent ones again, and asks nothing: a small note says what it learned, learned terms are listed under each context's terms with Remove, and the dictation that taught one offers Undo. Learn from your edits is on wherever Notice edits after pasting is, with its own switch in Settings → Privacy & Resources. A Learning page in the sidebar shows recent activity: what was learned and where, what was held back (with Learn Anyway), what was left out and why, and the terms it won't learn again (with Allow Again).
  • Armenian on the Mac is recognized by the speech model itself, taught Armenian by a 0.1 GB adapter that replaces the 0.9 GB Whisper model, whose files Owscribe removes: about half the errors on everyday and colloquial speech, and nothing changes for your other languages. A dictation that gets stuck repeating itself now stops within seconds.
  • When a newer version of a model you have is available, Owscribe offers it in the menu bar and in Settings, with its size, and downloads it only when you choose Update. The version you have keeps working until the new one is in place, and is then deleted.
  • Long Armenian dictations come out whole. Owscribe now splits them at pauses, skips stretches with no speech, and removes a word or phrase the model got stuck repeating, which could fill the rest of a dictation longer than 30 seconds.
  • A dictation no longer loses its end when the speech model briefly thinks you've finished: an ending it isn't sure of is passed over, and what follows is kept when the model then ends for certain.
  • A long dictation no longer comes back as a word or two when the speech model stops right after its opening: Owscribe notices there are far too few words for the recording and recognizes it again in parts cut at pauses.
  • A recording with no speech in it, such as a shortcut pressed by accident, ends as No speech recognized. The speech model used to write something for it anyway, such as “Okay.” or the name of the app and a term from your vocabulary, and Owscribe pasted that.
  • Choose the microphone in Settings → Recording. A microphone that stops responding can no longer freeze Owscribe: starting gives up after two seconds and says so, and the microphone test and calibration stop with a note when no sound arrives.
  • Pasting gives your clipboard back as soon as the app has taken the text, instead of after a fixed wait, and transcripts no longer land in clipboard managers' history. A password manager's copied password is never put back.
  • Owscribe no longer turns on VoiceOver's accessibility mode in the apps you switch to, which could slow down Chrome and apps built like it. It turns it on only in an app where it couldn't find the text field and that has no other way to show it, such as Codex, so pasting there works again.
  • Remember where you dictate keeps each dictation's app, window title and address until you delete the dictation or turn it off, instead of for 30 days. Owscribe deletes nothing on a schedule. Remembered web addresses also drop sign-in and share tokens from their path.
  • Use text near the cursor starts off on new installs, and edits after pasting are never watched in terminals or password managers.
  • When something fails, Owscribe says what and where: a model that didn't load, Dictations that couldn't be opened or cleared completely, or contexts it couldn't read.
  • Show Recording in Finder, in a dictation's ⋯ menu and when you right-click it in the list, takes you to its audio file.
  • In a narrow window, Dictations shows the list, and clicking a dictation opens it, with Back (or Esc) to return.
  • Quitting no longer waits indefinitely on a download, a model load or a transcription; it waits a few seconds at most.

0.11.0 · 2026-09-30

  • Contexts replace Projects. Owscribe recognizes what you're writing (an email, a chat message, a note, instructions for an AI agent, a commit message, and more) from the app and the web page, the project it's about, and the chat you're in, with no setup; each shows in the recording panel (with its icon, and a question mark when it's a guess) and Dictations, which also says which settings they changed, and Retry applies them again. Your projects carry over. Choose what you're writing or a project from the menu bar.
  • Turn on Pick what you're writing (Settings → Privacy & Resources) to let the local language model decide when the app alone doesn't tell, such as in a terminal or an unfamiliar app.
  • Turn on Remember where you dictate to have Owscribe add a chat when you dictate into one, and a project when an app shows the same one twice. It keeps each dictation's app, window title and address for 30 days, and forgets them when you turn it off.
  • In a chat, dictation starts in the chat's language. Owscribe remembers the language last recognized in each chat (while it remembers where you dictate), and with Read a conversation's language turned on it reads the last few messages on screen in Messages, and in Slack (in Chrome or its app), and Discord and WhatsApp in Chrome, to tell. Armenian from a chat is used on its own; other languages come first among yours. Set a chat's language, or forget the one it learned, on its page in Contexts.
  • Each kind of writing can set its own cleanup (Off, Simple or Local model) and add instructions for the local model. Instructions for an AI agent ship with instructions that keep paths, identifiers and commands as code spells them and number spoken steps. Search queries, terminal commands and task titles leave off a final period, and a commit message gets its summary on its own line.
  • Turn on Lay out emails or Lay out notes on those pages to have the local language model put a greeting and sign-off on their own lines and separate paragraphs, or turn items you list into a list. Owscribe checks that no word was added, dropped or changed, and pastes the cleaned text otherwise. It's off by default, since a dictated request to format something may be laid out instead of kept as said.
  • Owscribe keeps your corrections with each dictation, while it remembers where you dictate, to learn from later: Retry As another language, Change Writing (new in Dictations, which transcribes the recording again as another kind of writing), keeping the original over the cleaned text, and choosing a context in the menu bar where another was detected. Turn on Notice edits after pasting to also keep how you edited pasted text within a minute. Dictations lists them.
  • Setup and Settings → Privacy & Resources list everything Owscribe may read from other apps and the memory and storage it may use beyond dictating, each with its own switch, what it does, and what turning it off changes. Your current choices carry over; everything new starts off. Detect the project is now Read the window you dictate into, and Storage & Privacy is now Storage.
  • Cleanup's choice and instructions move to Contexts → Everywhere; Settings → Recording keeps the local language model's download and whether it stays loaded.

0.10.0 · 2026-09-30

  • Dictate in other languages. Choose yours in Setup or Settings → Recording, from the 30 the speech model supports; a new install starts with your Mac’s languages, and English stays as it was for everyone else.
  • With several languages, Owscribe recognizes which one you speak in each dictation, only ever among yours. Pin one from the menu bar’s new Language menu, or give a project its own language.
  • Dictate in Armenian, with a speech model of its own (a one-time 0.9 GB download offered when you add Armenian). Owscribe uses it when you pin Armenian from the menu bar, a project sets it, or it’s your only language. English terms from your vocabulary are spelled as you typed them, even when spoken inside Armenian.
  • The recording panel and Dictations show each dictation’s language. If it was wrong, Retry as another one of yours.
  • Settings → Recording → Cleanup: Off, Simple, or Local model, replacing Verbatim and Clean. Your choice carries over.
  • Simple cleanup follows each language: filler words are removed in English, German, French, Spanish, Italian, Portuguese, Russian, Japanese and Chinese, and left as spoken in the others. Words that are fillers in one language stay in another, such as the German “um”.
  • Local model cleanup runs a language model on your Mac (a one-time 2.3 GB download) that also removes stutters and false starts and applies spoken corrections, following instructions you can edit. Choose whether it stays in memory or loads when you dictate. It has been tested only in English, so dictations in other languages are pasted as recognized, and Dictations says why.
  • With text near the cursor turned on, pasted text fits what’s beside it whatever the cleanup setting. Chinese, Japanese and Thai text is pasted without extra spaces.
  • Dictations switches a cleaned-up dictation between the cleaned and the original wording; Copy copies the one shown, and search finds either.
  • Owscribe holds less memory between dictations, and when your Mac runs short of memory it frees the speech model and loads it again as you next dictate.

0.9.2 · 2026-09-29

  • Choose how the shortcut starts dictation: hold it, press it to start and again to finish, or both — tap to keep dictating, hold to dictate until you release. Esc cancels a dictation that continues without the shortcut.
  • A single modifier key can be the shortcut, such as Right ⌥ or Right ⌘; the key on the other side keeps working as usual. Fn on its own isn't allowed, and Owscribe explains why.
  • Clean transcripts: filler words like “um” and “uh” are removed before pasting. Choose Verbatim in Settings → Recording to paste exactly what was recognized.
  • With text near the cursor turned on, pasted text fits what’s beside it: a space where words would touch, a lowercase start mid-sentence, and no second full stop.
  • Dictations keeps the original wording of a cleaned transcript; Copy Original is in its ⋯ menu.
  • Setup shows how far the speech model’s download has got, and you can choose a shortcut while it downloads.

0.9.1 · 2026-09-28

  • Help menu: the website, release notes, privacy statement, and a way to contact support.
  • Settings → Storage & Privacy links to the privacy statement.
  • Update notes like these now appear when Owscribe offers an update.

0.9.0 · 2026-09-28

First public beta. English only; needs a Mac with Apple silicon and macOS 26 or later.

  • Hold a shortcut, speak, and release: Owscribe transcribes on your Mac and pastes into the field you were in, or copies the text when it can't paste.
  • Recognition runs entirely on your Mac, with the Qwen3-ASR speech model, downloaded once (about 2.5 GB).
  • Projects give names and terms their right spelling, applied when Owscribe detects the project from your window or when you choose it from the menu bar.
  • Works without the Accessibility permission, copying instead of pasting.
  • Dictations keeps every recording and transcript on your Mac, with playback, retry, and search, until you delete them.
  • Updates arrive through the app itself.