Every version, and what
each one changed.

Patch

1.14.1 · 29 September 2026

Choose who reads, and how fast. Speak Translation now offers two voices: the best one installed for the translation's language, as before, or VoiceOver's own. And four speeds: Slow, Normal, Fast and Very Fast.

• VoiceOver's voice, the app's timing. VoiceOver Voice reads with VoiceOver's settings, in VoiceOver's language whatever the translation's, and the original still steps back while it speaks. With VoiceOver off, the Mac's own system voice reads.
• Slow, Fast and Very Fast read at about 0.8, 1.3 and 1.9 times the usual pace; with VoiceOver Voice, the usual pace is VoiceOver's. Nothing is skipped at any speed.

Minor

1.13.0 · 26 September 2026

Hear the translation. A new Speak Translation menu reads the translation aloud as each sentence settles, in the best voice installed for its language: on whenever VoiceOver is, or Always. Nothing is skipped and nothing is sped up, and ⌥⌘. stops it or reads the last line again.

• The original steps back. While the translation is read, the original drops to a quarter of its volume and comes back after.
• Made for VoiceOver. VoiceOver reads the caption box and presses its button, and the offer to save the transcript is spoken as it appears. VoiceOver's own speech is never captioned or translated.
• Talk over it. With the microphone and the Mac's speakers, the translation waits for a pause, and anyone who talks while it reads is still captioned, with the translation's own voice taken out of what the microphone hears.
• Smaller things. The offer to save the transcript comes after a minute or more of speech and Escape takes it away; Save Transcript… is in the menu whenever there is something to save. AirPods' microphone no longer crashes the app.

Minor

1.11.0 · 25 September 2026

Subtitles in your language. The menu, Settings, the welcome window, every other window and the status line are now in the sixteen languages the app transcribes, each written for its language rather than translated word for word. The app follows your Mac's language.

• A Language pane in Settings. Choose System Language or any of the sixteen. The menu and any open window switch at once, with no restart, and the choice is the one System Settings keeps for each app.
• The welcome window's demo speaks it too, as the one on this site does in every language, and it can be paused and restarted.
• Clearer words in Settings. The Pointer Reveal slider is now called Fade, and the ⌥ stack's settings say what they do in either color theme.

Minor

1.10.0 · 23 September 2026

The new speaker's first words go in their own box. With New box on speaker change turned on in Settings, the new box used to open a second or two after someone else started talking, so their first words ended the other person's box. The break now goes where their voice began.

• Placed where the voices change. Speaker detection (diarization) only knows someone else is talking a second or two late, so the change now carries the moment the new voice began, and the box breaks at the end of a sentence or the pause between the two voices, whichever is nearest.
• Words already shown move with it. Words already on screen go to the new box. In the ⌥ stack, words that closed with the other speaker's box are taken back out of it and join the box on screen, or become a box of their own if the box had faded.
• A cough is not a new speaker. A voice has to hold for half a second before it counts, so a cough or a quick "mm" no longer breaks the page. With translation on, the translated box breaks at the same point once its words there have settled.
• The ⌥ stack follows you across Spaces. Stop the audio on one desktop, move to another and hold ⌥: the stack now opens there, as the live box always did, rather than on the desktop where it was last shown.

Patch

1.8.4 · 21 September 2026

The name on the box follows the words in it. Under all system audio, when a second app starts playing, the app listens to it for three seconds, as it has since 1.7.1, and now asks whether there is a voice in what it hears and, under the Multilingual model, transcribes that moment on its own and holds the words against the ones on screen. An app heard saying the words on screen takes the box's name and icon; one heard playing music, with or without lyrics, or any voice other than the one being captioned, does not, however long it plays. A playlist started during a call leaves the call's name on its words.

• Listened to again every twenty seconds, in case that changes: a call app heard playing hold music is named once someone picks up and the words on screen are its own.
• When the named app stops, the box falls to an app heard speaking before one heard playing sound, and to that before one heard silent. An app heard speaking also keeps the box against one that could not be listened to, which used to take it after thirty seconds.
• The captions themselves are untouched. The check runs a second recognizer over the same loaded model, on the app's own audio, for a fraction of a second on the Neural Engine per listen. Under the other models the check stops at a voice, and with Skip non-speech off in Settings, or while the model is still loading, sound counts as words as before.

Patch

1.8.2 · 20 September 2026

Audio Borealis. Color along the bottom of the box that rises and falls with the sound. Soft lobes fan out from the edge, each following a band of the voice, and over them one translucent curve per band, lows to highs, lifts as its band is heard, so a vowel and a sibilant move different hills. Silence is nothing; a voice makes it pulse with the syllables. It follows whatever the app listens to, an app's audio or the microphone alike, scaled to its own peaks so a quiet voice across a desk fills the same range a film's soundtrack does. On by default, in the new Audio Borealis submenu of the status menu: Off, or a look (Rainbow, Northern Lights, Autumn, White Haze) and a strength (Strong, Medium, Subtle). Rainbow and Medium unless you choose otherwise.

• The welcome window's demo glows the same way, and follows the look and strength you pick. So does the demo on this site, on a voice it makes up.
• The Audio Borealis sound glow is open source for the web: hello-mat.com/design-engineering/audio-borealis.

Minor

1.8.0 · 19 September 2026

Listen To now has the microphone. The boxes wear a white mic on a red tile as their source, and the choice is remembered across launches. macOS asks for the microphone the first time it is chosen and not before, never at launch, and a refusal opens System Settings at the microphone list. A microphone is mono and not always 48 kHz, where the taps are 48 kHz stereo, so the audio core is rebuilt for whatever format the input has, on the way in and on the way back.

• Show Both Languages, in Translate To. This feature enables holding multilingual conversations live with both speakers seeing their own language. With it on the box carries two paragraphs: the Translate To language on top as the caption, the other language under it in a smaller, dimmer run. When the target language itself is spoken the pair reverses, the speech as the caption and the translation under it, with its unsettled tail dimmed as the caption's is and following Translation Timing the same way. The same language never appears twice: the lower paragraph is only ever the words the translation was made from, never the live transcript, and a sentence the translator returned unchanged is left out. A page turns when either paragraph fills, and is held long enough to read the slower of the two. Each box in the ⌥ stack keeps both texts as it showed them. ⌃ does nothing while both are shown: there is no original left to peek at.
• Translation works in turns. The recognizer keeps one transcript across a conversation, and finishing a sentence used to rewind and re-translate all of it. Now a pause or an endpoint closes a turn, words arriving while it settles wait for the next, and changing the target re-translates only the current turn. A tail dimmed as unsettled no longer stays dimmed for good, and a request the translator leaves unanswered for ten seconds is sent again.
• The welcome window's demo listens where the app does and shows both languages when the app does, from the moment it opens and as either is changed, with the menu bar's orange microphone pill up while it listens. Its boxes sit at the app's line height, as this site's now do.
• The question about checking for updates automatically is asked as the welcome window closes, in that window, rather than at the second launch. A first-time user closed the welcome and heard nothing about updates until the next day's relaunch. It is still the one question, asked once.

Patch

1.7.1 · 14 September 2026

A notification sound no longer puts its app's name on the box. Core Audio reports an app as playing for as long as it holds an output stream open, not for as long as it makes sound, and a notification holds one for seconds after the ding: about two in a native app, ten in a Chromium-based one such as Slack or Discord, and for a whole burst of messages the stream never closes at all. So a message arriving during a video renamed the box after Slack for ten seconds, and a busy channel kept it there. The rule from 1.7.0, that a newcomer had to be playing for two polls running, was written for a ding that lasts half a second, which none does.

• A newcomer is listened to before it can take the box. Two seconds after an app starts holding a stream, by which time its own ding is over, the app taps just that app for three seconds and counts how much of it clears a floor; only sustained sound earns the name. A ding reads as silence and changes nothing, a burst of dings likewise, and an app heard to be silent is listened to again twenty seconds later in case the burst became a call. A call that starts while a video plays is named within about six seconds.
• A backstop, and a fallback. An app nobody managed to listen to takes over after thirty seconds; one heard to be silent never does, however long it holds its stream. When the named app stops, the box falls to the best of the rest, never to one heard to be silent.

Patch

1.6.5 · 12 September 2026

Translation starts the moment it is turned on. It used to wait for the recognizer to name the language it was hearing, which the multilingual model does at the start of a sentence rather than on the first words, so captions that began mid-sentence, or a target picked in that stretch, stayed untranslated until the next full stop. The language is now read off the transcript itself until the model names one, and turning translation on or changing its target starts the recognizer afresh, so the next words open a new box in the new language instead of finishing the old one.

• A shorter menu. Skip non-speech and New box on speaker change live only in Settings now, under Models, where they already had switches: each one reloads the engine, which is not a thing to flip from a menu while watching. Reset Overlay Position sits in a group of its own, apart from the toggles above it.
• Acknowledgements is a button in the About window, beside a new Changelog button that opens this page. It opens a window of the app's own rather than a text file in another app: the third-party notices, reflowed into paragraphs and set in the app's type, in a soft box that scrolls.
• The trial ending puts the same red badge on the menu bar icon that a waiting update does, so it is seen without opening the menu. The badge is never dimmed: pausing fades the icon, and used to fade the badge with it. An update and an ended trial at once show a 2.
• The update window's release notes box shows in light mode now. Its gray was too faint to read as a well. It also takes its color when drawn rather than when the window is built, so a window open across a change of appearance gets the right one.

Patch

1.6.3 · 9 September 2026

A license key can be pasted into the license window. ⌘V did nothing there, and neither did ⌘C, ⌘X, ⌘A or ⌘Z: those keys reach a text field only through an Edit menu, and an app with no menu bar had none. It has one now, which nothing ever shows, and the Clear after field in Settings takes a paste for the same reason.

Patch

1.6.1 · 8 September 2026

The ⌥ rewind stack no longer shifts sideways when a box arrives that is wider than the ones already up, or when the widest one leaves. The stack is as wide as its widest box, so its edge moves at those moments; what moved with it, for a tenth of a second, was every box in it. The stack now follows the live box by its center, and each box is placed from that center, so the stack's width can change under them without a box moving by a point.

Minor

1.5.0 · 7 September 2026

The app updates itself. Check for Updates… sits at the top of the menu, under Pause. It asks this site for a newer version and installs it in place, same location and same signature, so the audio permission survives an update the way it survives a rebuild. Copies older than this one have no updater; they are updated by downloading again, once.

• It asks first. On the second launch, once, whether it may check on its own once a day. The request carries the app's version and nothing else, no audio, no captions, nothing about you, and the answer is a checkbox in the menu afterwards. Until it is yes, the app makes no request at all.
• Never over a call. A check that finds something opens the update window at launch, or once nothing has been playing for a couple of minutes. While something is playing it does not: a red badge with a 1 appears on the icon, the menu offers Update to …, and the window comes when you choose it. A critical fix opens the window regardless, and cannot be skipped.
• One window of the app's own. The notes set as text, one bar for the download, Install and Relaunch when it is ready, in the style of the About and Welcome windows rather than the updater's stock dialogs.

Patch

1.4.2 · 6 September 2026

The welcome window's demo searches the ⌥ rewind stack, as the app does. A vendored copy without the search had nothing to attach it to, and stopped on load.

• The settings preview's screen is the demo's again. Its colored desktop, a menu bar with this app's glyph and the clock in it, and a call in front with the ring handed round whoever is talking. The ⌥ rewind stack stops at the menu bar, as on a real screen, and the preview is a little taller.

Minor

1.4.0 · 6 September 2026

The ⌥ rewind stack can be searched. A small field sits at the edge of the stack, against the live box, drawn as one more box. Click it, or press ⌥F while the stack is up, and type: the stack narrows to the boxes that contain the text, ignoring case and accents, lights the matches and scrolls to the newest one. While the field has the keyboard the stack stays up with ⌥ released. Escape, or a click anywhere outside it, lets it go again.

• It keeps every box by default. The Keep slider in Settings runs one stop further, to ∞, and that is where a fresh install now sits: the stack holds every box that closes, up to two thousand, in place of the last fifteen. How long a silence takes to forget the stack is the setting it always was, and the two are separate.
• The stack follows the live box on a spring. The box grows a line at a time as a sentence wraps, and the stack used to jump the line with it. Now it catches up over a tenth of a second, while the box itself stays exactly where it was.
• Deep stacks no longer hitch. Every box was measured again each time the stack was rebuilt, which became a pause at the depths the slider now allows. The sizes are kept between rebuilds.

Patch

1.3.4 · 3 September 2026

The welcome window's demo caught up with the site's. It plays the same three scenes the home page does, now under the same menu bar: the Apple mark, the app in front, the first menus, and a desktop with some color in it. The podcast scene's waveform fills its row again too, which the app's copy of the demo had been missing.

• The screen's corners are tighter in the welcome window than on the site, so they no longer clip the Apple mark at one end of the menu bar and the clock at the other.

Patch

1.3.2 · 3 September 2026

Text stops coming back after a pause. A box would fade, the next one opened with words said before it, and from then on it filled without ever turning over. Three faults, all of them the same mistake in different places: treating words still being spoken as though they had settled. Whatever a faded box was still owed is now dropped along with it.

• A sentence nobody finishes no longer fills the box for ever. Text that has not settled cannot be paged at all: it is retranslated whole on every update, so its words move each time and a break put inside them does not stay where it was put. It now settles on length as well as on age, after which the box turns over like any other.
• Pages break at the edge of the unsettled text. A break made inside it came undone on the next update, which is what put the same few words in three boxes in a row while nothing new was being said.

Minor

1.3.0 · 31 August 2026

Captions can be translated as they are spoken. Pick a target in Translate To and any of the sixteen languages arrives in any other, through Apple's own translator, on the machine. Nothing is uploaded, which is the property the whole app is built around. How it works.

• Every pair, not only the ones with English in them. All 240 orderings of the sixteen. Speech models that translate as they transcribe go to or from English only, which would have left most of these unreachable.
• Two timings, because the question is what to show mid-sentence. Live, Then Settle dims text that may still change and leaves it alone once it has stopped. Always Live retranslates everything on every update: no lag, and nothing on screen is final.
• Hold ⌃ to see the original. The language being spoken, live box and ⌥ rewind stack together, for as long as the key is down. Translation carries on underneath, so letting go shows current text rather than a snapshot.
• A ghost mode for screen sharing. Show Overlay In Screen Share / Capture is on by default. Uncheck it and the overlay leaves anything capturing through macOS: calls, recordings and screenshots alike.
• Interrupted model downloads repair themselves. A download that stopped partway used to leave a bundle CoreML refuses, and nothing ever retried it, so every launch after that failed with nothing on screen to say why. Those are now spotted and refetched, only the broken ones, and a recognizer that cannot load says so in the menu.
• Translation needs macOS 15, which is where Apple's translator arrives. The rest of the app still runs on 14.2, where those two rows are simply not in the menu.

Patch

1.1.4 · 22 August 2026

A settings window that previews itself. Half of what the sliders govern describes something not currently on screen: boxes that have already finished, the hole that follows the pointer. So the pane got a small overlay of its own to demonstrate on. Touch a control and the preview says what that control does, with your current value in the sentence.

• Windows that fit. Settings and the welcome screen scroll on a short display instead of running off the bottom of it, where there is no title bar to drag them back by.
• The overlay follows a resolution change rather than waiting for the next word to notice one.

Patch

1.1.2 · 19 August 2026

The welcome screen now says what ⇧ and ⌥ do and what the pointer does, one line each, with the right key lighting up for each one.

Minor

1.1.0 · 19 August 2026

Hold ⌥ to rewind, and the last few boxes come back. The overlay pages like broadcast subtitles. It fills, clears and starts again, so a sentence you glanced away from was simply gone. The closed pages now stack above the live box, and scroll when there are more of them than there is room for.

• A settings window behind ⌘,: lines per box, opacity, how long the rewind lives, and how far back ⌥ reaches.

Patch

1.0.2 · 16 August 2026

A welcome screen for the first run. First launch fetches about 600 MB before a single caption can appear, and until now the whole first impression was a menu bar icon with a dot on it. There is also an About window saying what version you have and where it came from.

Major

1.0 · 15 August 2026

The first release. Live captions for anything your Mac plays, drawn as an always-on-top overlay that stays put while you switch apps, with the transcription running on the Apple Neural Engine and nothing leaving the machine.

• A streaming model, not a 30-second window. Captions arrive as the sentence is being said rather than after it.
• Music never reaches the recognizer, so a backing track does not become invented lyrics.
• Language first, model second. The menu bar leads with the language and says what each choice costs in megabytes before you click it.