We started 1.6.0 from what people complain about most in dictation and meeting apps. A two-minute take gone because of one stray key. Text pasted into the wrong window, or a clipboard manager full of dictations. An AI "cleanup" that answered the question instead of typing it. And meeting notetakers nobody on the call could trust: recordings lost when the app quit, notes that didn't survive a crash, AI answers quoting things nobody said. Some of these were ours: until now, a stray Esc threw a Vext take away for good.

1.6.0 is built around those complaints, and it adds three headline features: Voice Edit, Apple Intelligence as an on-device AI provider, and Ask your meetings with quoted sources. It follows 1.5.1 and is a free update for every Vext 1.x license — a license covers the whole major version.

Dictation you can get back

  • Undo a cancel. Esc or ✕ now shows Discarded with Undo for six seconds. Up to ten minutes after the cancel, Recover Discarded Take in the menu bar brings the take back and pastes it into the app you dictated in.
  • No pasted noise. A take that's only a cough or a breath says Didn't catch that and pastes nothing; a whisper still counts as speech. Failed takes keep their audio for Retry in History, and voice notes now do too.
  • Mic changes mid-take. AirPods connecting or a dock unplugged no longer cost you the rest of a dictation, note or meeting. The pill says "Now using ‹mic›", or "Can't hear you — ‹mic›" when the mic goes silent.
  • Pause and a safe stop. Pause a hands-free take or a voice note from the pill, the menu bar or the Pause Dictation or Note shortcut; the mic closes until you resume. Hands-free dictation now stops after one minute of silence by default, with a "Stopping in N s — speak to keep going" countdown first.
  • Quitting finishes the take. Quit mid-dictation and the pill says Finishing your dictation… while the take is saved to History. Quit Now skips the AI pass and quits within five seconds.

Text that lands where you meant

Vext now sends keys only to the app you dictated into. If that app can't come back to the front — you quit it, or focus moved elsewhere — nothing goes into whatever is in front. The text waits on the clipboard with a Copied — press ⌘V to paste toast and a Paste button.

  • A clean clipboard. Writes made only for a paste are marked transient, so clipboard managers skip them, and your previous clipboard comes back afterwards.
  • Context-aware text. Where the app lets Vext read the text before the cursor, continuing a sentence adds the missing space and lowercases the first word: "I think" plus a dictated "That works" becomes "I think that works". Names, "I" and your vocabulary keep their capitals. Short messages in Slack, Messages, WhatsApp, Telegram, Discord, Teams and Signal drop the final period.
  • Insert text by pasting, Typing it out for terminals and remote desktops that ignore ⌘V, or Copying to the clipboard only — globally in Settings → General, or per app. A dictated new line no longer sends a chat message or runs a terminal command.
  • Spoken punctuation and commands (off by default): "period", "new line", "scratch that". Code commands are off everywhere, terminals included, until you tick them for an app in Settings → Language & AI → Per-app styles; then "git checkout dash b main" becomes git checkout -b main.
  • Say "send" to submit (opt-in): end a hands-free dictation with "send" after a pause, and Vext presses Return. Push-to-talk never submits.
  • Paste Last Dictation and Copy Last Dictation in the menu bar rescue a paste that went astray.

Voice Edit

Select text anywhere, hold the Voice Edit shortcut, say how to change it — "make this more formal", "translate to German" — and let go. The selection is replaced. With nothing selected, say what to write and it appears at the cursor. Voice Edit has no shortcut until you set one in Settings → Keyboard Shortcuts, and it needs an AI model.

The rule we care about most: a voice edit checks the selection before it replaces anything. If the selection changed while the edit ran, or the app won't let Vext read it (the pill warns "Can't read the selection here · the result will be copied" up front), nothing is replaced. The result waits on the clipboard with a Paste button, and History keeps your instruction and the text Before the edit.

Apple Intelligence, and AI output that's checked

On macOS 26 with Apple Intelligence turned on, choose Apple Intelligence under Settings → Language & AI → AI Processing, or accept it when Vext offers it in place of a model download. Choosing Clean up in the welcome guide also picks it when no local model is downloaded. Styles, note cleanup, summaries, Ask and follow-ups run on Apple's on-device model with nothing to download. In a language Apple Intelligence doesn't support, dictation uses the local model if you've downloaded one; otherwise it's pasted as spoken and the pill says why.

In the built-in styles, AI output is now checked before it's pasted. A reply that answers your question instead of transcribing it, adds a preamble, switches language or gets cut off is discarded, and the pill says Pasted as spoken with the reason. Dictate "what is the capital of France" with Clean up and you get the question, not "Paris".

  • Paste original in the ⋯ menu of the Pasted chip inserts your words as spoken.
  • Rewrite as in History restyles a saved dictation or note without re-recording. The original is always kept.
  • The menu bar has a Dictation Style submenu, the pill shows the active style, and three presets join the list: Bug report, Implementation brief and Task.
  • Downloads on request. The local model downloads only when you press Download, which shows the size (3.44 GB for the default Gemma 3 4B), with progress and Cancel. Downloaded models lists what's on disk with Delete, and the model leaves memory after 10 idle minutes by default.
  • Long meetings are summarized in parts. A summary, answer or follow-up cut off at the length limit is written again with more room or reported as failed, never saved half-finished.
  • Summary style: seven styles, among them Standup, Sales call and Decision log, as a default or per meeting. Custom… takes your own instruction and keeps it for next time.
  • Separate translation. Translate dictation and notes to and Translate meetings to are now two settings. The meeting target also sets the language of summaries, answers and follow-ups.

Your language and your words

There's now one Dictation language setting for every engine, the first row under Settings → Audio & Transcription → Transcription for dictations and notes. Parakeet covers 25 European languages, and the Automatic option never sends audio to Apple.

A language Parakeet can't transcribe, such as Hindi or Thai, goes to Apple Speech. On macOS 26 that's SpeechAnalyzer on this Mac ("Hindi uses Apple Speech on this Mac."), for meetings too. On macOS 14 and 15 Apple's older recognizer sends audio to Apple's servers, and Settings says so before you dictate; meetings in those languages need macOS 26.

Settings → Vocabulary → Teach Vext your words adds a name, brand or acronym that Vext writes exactly as typed. With Boost recognition of my words on and the optional 102.8 MB Parakeet CTC 110M model downloaded, Parakeet also listens for those words in English dictation, voice notes and meetings. Boosting needs 3 or more letters, A to Z only; other words are still written as typed. A word is swapped in only where Parakeet was unsure and the audio fits yours better, so teaching "Vext" doesn't turn every "next" into "Vext".

We measured it on synthetic speech: test sentences read by five macOS say voices, with 23 taught words. Hard terms went from 11 of 65 right to 38 of 65. Ordinary words didn't get worse: 3 errors out of 310, before and after. That's a synthetic test, not a promise about your voice or your microphone. The model knows only English.

Correct and remember covers the rest: fix a word in a finished dictation from the pill or History and keep the fix as a rule. More on getting names right in fixing Mac dictation accuracy.

Meeting transcription that recovers from quits and crashes

Meetings got the most work, because a notetaker you can't trust isn't worth running.

Nothing gets lost. Quitting while a meeting records asks: Stop and save, Keep recording or Quit without saving; logging out saves it. Audio and notes are checkpointed about once per second, so after a crash, Recover Interrupted Meetings… in the menu bar brings back the transcript, notes and screenshots. A meeting with no speech still keeps its notes, and if some stretches failed, Transcribe missing parts fills only those and keeps speaker names and the summary.

You know what's being captured.

  • If System Audio Recording is blocked, the pill marks system audio Not captured with a button to fix it. Allow it mid-call and the other side starts recording without a restart. A mic that dies shows "Microphone stopped — Vext can't hear you".
  • On macOS 14.2 and later, the record prompt names the app actually using the microphone, browser calls included.
  • About 20 seconds after the call's app lets go of the mic, Vext asks Call ended — stop recording? It never stops on its own.
  • Dictating during a meeting doesn't lower the call's volume, and the dictated words stay out of the meeting transcript.

Tools while you talk.

  • Mark moment from the pill, the menu bar, the notes panel or a shortcut. Moments show in the transcript, the summary and exports.
  • The notes panel time-stamps every line you type, so a line jumps to its moment later. A Live transcript in the panel shows recent lines with speaker names, 10 seconds to 2 minutes behind.
  • Share consent notice pastes an editable notice into the call app's chat (you press Return) and copies it. The meeting records when it was shared and whether it was pasted or only copied.
  • Off the record pauses both streams until you resume, and the transcript shows "Paused for N min".
  • Vext marks its pill, notes panel and toasts to be left out of screen shares and recordings, and leaves them out of its own screenshots.

Optional, off until you turn them on. With Use Calendar for meeting titles and attendees (Settings → Audio & Transcription → Meetings), a recording made during an event with other people or a video link takes the event's title, attendees are suggested as speaker names, summaries get the agenda, and Vext can remind you to record a minute before. Keep meeting audio for playback adds a player, Play from here on transcript lines and Audio (.m4a) export, and deletes the audio after 30 days by default.

And Import Recording… turns audio and video files into meetings with speaker labels, like a recorded call.

Ask your meetings, search everything

  • Ask your meetings in the Meetings toolbar asks across a date range; Ask about this meeting saves its answer with the meeting. Every statement must cite a verbatim quote, statements without one are left out, and clicking a source opens the meeting at the quote.
  • Create follow-up drafts a Client follow-up, Engineering handoff or Action list linked to the transcript lines behind it.
  • A saved Action list fills Action items in the sidebar. Add open items to Reminders creates one reminder per item, never twice, and completing one in Reminders checks it off in Vext.
  • Search across dictations, notes and meetings is ranked by relevance, ignores accents, matches prefixes and opens at the matching segment.
  • Vext recognizes a voice in later meetings only after you tick Remember this voice on this Mac, and Forget voice removes it. Until then, a meeting keeps a voice sample of each speaker it didn't recognize, and the sample is deleted with the meeting. Change speaker… moves a single line to someone else.

Privacy & Data

Settings has a new Privacy & Data pane:

  • Local only (off by default) blocks cloud transcription and AI, including requests already under way. A server on this Mac, such as Ollama on localhost, still works.
  • Keep history deletes old dictations, notes and meetings on a schedule you set per type, and tells you how many items will go first.
  • Permissions and Storage show what Vext can access and what it keeps on disk, with sizes, and Delete for audio, recovery data and AI models.
  • Copy Diagnostics copies details for support with no transcripts, names, paths or keys.

Deleted items are now zeroed out in the database file and erased from the search index, instead of lingering in free space.

The MCP server gained library search, cross-meeting Ask and action items, and it can tighten privacy settings but never loosen them: an agent can't turn off Local only, keep history longer, keep meeting audio, or turn on the calendar. One limit to know: Local only covers Vext's own providers, and an MCP client you connect can still read your library, so connect only apps you trust.

Smaller things

  • First run. Only Microphone and Accessibility are required; Vext no longer asks for Input Monitoring. The welcome guide checks your mic, asks which language you'll speak, and takes a practice dictation that doesn't count against your 100 free dictations.
  • Notifications instead of pop-up windows. "Meeting saved", then "Summary ready", as native notifications that open the meeting.
  • Menu bar status. It shows when the speech model is downloading and a meeting is saving, and it blinks red for as long as a meeting records.
  • vext:// links for Stream Deck, Raycast, Alfred and Shortcuts start dictation or meetings, mark a moment or paste the last dictation. They're off until you turn them on in Settings → General → Links.
  • A compact pill. After a paste the pill shrinks to a small Pasted chip, and Settings → General → Dictation status → Top center moves it away from chat boxes.
  • Stats and export. The Dictations tab shows words, time saved and your streak, and Save every finished meeting as Markdown writes meetings into a folder you choose.
  • A new meeting shortcut for new installs: Left Control + Left Command. Existing installs keep Fn.

Update

brew upgrade muvon/tap/vext

Or download Vext 1.6.0 directly. It's free for every Vext 1.x license.

Dictations, notes, meetings, speakers, vocabulary, styles and shortcuts carry over. The first launch rebuilds the search index in the background, which can take a few seconds on a large library.

A few defaults change for everyone, existing installs included. Hands-free dictation stops after one minute of silence; set Stop hands-free dictation after silence in Settings → Keyboard Shortcuts to Never to keep the old behavior. Short messages in chat apps drop the final period; untick Drop the final period in short messages for an app under Per-app styles. And a saved meeting shows a notification instead of opening Vext's window, unless you turn on Open Vext when a meeting is saved in Settings → Audio & Transcription → Meetings. Voice Edit, Pause, Paste Last, Mark Meeting Moment and Off the Record have no shortcut until you set one, and every opt-in feature above stays off until you turn it on.

New to Vext? It includes 100 dictations, 50 notes and 10 meetings free, then it's $49 once, with every 1.x update included.

Everything still runs on your Mac unless you choose otherwise: no account, no subscription.

Download Vext 1.6.0