Changelog

1.1.0-beta.4

Mac

• Voices in a room are named while the recording is still going. With the microphone records a room on, each turn in the live transcript gets a speaker — Speaker 1, Speaker 2, or a remembered name — instead of everything being "Me" until the finished document arrived. It is a live guess and not the record: a short interjection may take the label of whoever was already talking, two similar voices may come out as more speakers than there are, and nothing is revised once written. The finished document is unchanged and remains the version to trust. On a call nothing changes at all — the other side still arrives on its own track, which is better evidence than any clustering.
• A long stretch from one speaker breaks into paragraphs instead of growing into a single block. Each paragraph carries the time it began, so clicking a timestamp lands on the words you are reading rather than at the top of a ten-minute run.
• The transport icon is drawn at the size of its neighbours. It was about two-thirds the size of the other menu-bar icons, which read as "small" without being something you could name.

1.1.0-beta.3

Mac

• The microphone works in the downloaded build. beta.2's notarized signature did not carry the audio-input entitlement, and under the hardened runtime that refuses the microphone outright — no permission prompt, no error, no audio. Builds made on the developer's own machine never showed it, which is how it shipped. If beta.2 recorded nothing for you, this is why.
• Releases are now built, signed and notarized on GitHub rather than on a laptop. Nothing visible changes; it is what stops the fault above from going out the same way twice.

1.1.0-beta.2

Mac

• Restarting Transcripts while it is recording no longer ends the recording. It picks the same one back up: the piece already captured is set aside, capture resumes within a second or two, and stopping reassembles every piece into one recording on a single timeline. Quit it, crash it, or rebuild it mid-session — the evening still comes out as one take. Only the gap has to be short (about ten minutes); a marker left by a crash days ago is filed, not resumed.
• The menu says "Resumed after relaunch" while that is in effect, so it is never something you have to take on faith.
• The menu-bar icon now follows tape-transport convention by default: a square when stopped, a hollow record light while auto-record is watching, and a filled disc while rolling that flashes from red towards white and swells as the room gets louder. Three shapes rather than one shape in three shades — which is the difference between reading the state and squinting at it — and a disc that brightens when somebody speaks, so a dead microphone looks different from a quiet room. The Transcripts mark, the waveform and the microphone are all still there in Settings, unchanged; an existing install keeps whichever it was already set to.
• Live text now appears as it is spoken rather than when the recogniser decides a phrase has ended. The overlay's pill shows the words mid-sentence, and the live transcript file grows a Being said now section at the bottom carrying the same not-yet-final text. The transcript itself is untouched by this: it still only ever contains finalised turns, because a record that revises itself is not a record.
• Recordings of a room are now split by voice. A microphone with five people in front of it used to come out labelled entirely as you, because the whole attribution path assumed a call — mic is you, system audio is everyone else. Turn on "My microphone records a room" in Settings and the microphone track is diarized instead, each voice getting a number or a remembered name. Detected calls are unaffected.
• Voices are matched against every sample remembered for a person rather than the average of them, and one person can now own several voices in a room recording. Both exist for the same reason: someone running a character at a table does not sound like themselves, and averaging the two produced a voiceprint matching neither.
• A new overlay: a small glass panel floating over your call. At rest it shows the last thing said, straight from the transcript with no summarizing in the way. Hover it and it opens into three lanes — the last conclusion, facts and figures, and the last question with its answer. The live transcript was already writing all of this down, but reading it meant looking away from the meeting, which is the one thing you cannot do during a meeting.
• Answers come only from earlier in the same call or from the Markdown notes under your transcripts folder, and every one says which, and when. The model phrases an answer out of text that was found first; it is never the source. A question nothing supports is shown unanswered rather than filled in, and an answer that does not hold up against the passage it came from is discarded before it reaches the screen — on a live call an invented answer is worse than an empty panel. The same rule governs the conclusion and the figures.
• Off by default. Settings ▸ General ▸ Overlay, with a separate switch for whether your notes are searched. Everything stays on your Mac.
• With no language model available the overlay quotes instead of phrasing: it shows the best-matching passage verbatim with its source, and leaves the conclusion and figures lanes empty rather than guessing at them.

1.1.0-beta.1

iPhone and iPad

• Transcripts filed by the Mac can be renamed, archived and deleted from the phone. Only recordings still waiting to be picked up could be, which left the finished article — most of the library, and all of what a phone is usually doing — with no way to fix a title the model got wrong. Swipe one way to rename, the other to archive; delete is behind a long press and asks first. Renaming changes the title inside the transcript, so every device sees it, and leaves the filename alone so nothing that links to the note breaks.
• Archiving moves a transcript and its audio into an Archive folder beside your transcripts, keeping the folder it was filed into. Nothing is deleted, every device sees the same archive, and unarchiving puts it back exactly where it was. Deleting moves both to the Trash, where Files can still recover them.

Mac

• The transcript reader can play the recording it was made from. It could show you what the model heard and give you no way to hear it, which is worst on the lines it got wrong — a garbled sentence is visibly wrong and unrecoverable without the tape.

Both

• Transcripts carry the time each speaker started, and clicking one jumps the audio there. The timings were always being computed and thrown away one line before the document was written. Turns now read Me: [12:04] …, in the finished document and in the live mid-call transcript an assistant reads during a meeting.
• Recordings made before this update have no timestamps and read exactly as they did. New ones get them; there is nothing to migrate, and nothing to turn on.

1.0.6

• Settings stops lagging when you click between its tabs. Every render was enumerating audio devices, resolving the chosen input again for each row, and asking the system about the login item — and because the tab bar rebuilds all five tabs, one click paid for all of it. The device enumeration also wrote the config file to disk mid-render, which invalidated the pane it was drawing.

1.0.5

• The iPhone and iPad icon is the violet tile the Mac already wore. The white one had no edge against the App Store's white page or a light home screen.
• Sessions can be made in Settings ▸ Sorting ▸ Sessions instead of only by hand-editing routing.json — name, folder, when it ends, and what runs when it does. Starting one from Shortcuts always worked; the picker was just empty unless you had edited the file.
• A title containing # or _ survives being read. "C# to F# Migration" was being filed as "C to F Migration" — in the transcript, its filename and the Recordings list — because unwrapping the model's markdown reached inside the title instead of stopping at its ends.
• A transcript mirrored into an Obsidian vault no longer offers to play audio that was deliberately left behind in the knowledge root.
• A transcript that reads but has no body says so, instead of showing an empty page.

1.0.3

• The menu-bar mark is one line of text over a waveform, not three stacked rows. The fuller mark was a picture of a document, and at 16 points a picture of a document is a smudge — beside wifi and battery it was visibly busier than every neighbour. It sits lower too: a solid line over a mostly-empty waveform puts the optical centre above the geometric one, which read as the icon floating above its neighbours.
• A shared recording could play back truncated. Its audio was copied straight to its final name in the cache, so a copy interrupted partway left a short file that every later play accepted as complete.
• A transcript that hadn't finished syncing could appear in the library as a bare filename with no date and no audio, and never ask to be downloaded.

1.0.2

• Recordings are named after what was said in them again. The summarizer is asked for a title and reliably gives one, but wrapped in whatever markdown it favours — and a decorated TITLE: … was not being read, so a recording kept the name of the window it was captured from. The two apps then disagreed about what to fall back to: the same call was “with Candidate” on the Mac and “Microsoft Teams — Aug 24” on the iPad, when it was an interview with a name. scripts/repair-titles.py puts the recovered titles into recordings made before the fix.
• The Obsidian mirror refuses to run when it would write over the copy it is mirroring — pointing it at the knowledge root used to blank the audio_file the filed transcript needs to find its own audio.

1.0.1

The release the first iPad taught us to make.

Obsidian

• Transcripts finds your vault and files a copy of every finished transcript into it, in the same folder Sorting routed it to — on by default if you use Obsidian, and a Settings row away from being a different vault or none. It is a copy rather than a move, because the iCloud folder is what your phone and iPad read their library from and a vault carried by Obsidian Sync is not something they can see. Markdown only: the audio stays in the knowledge root rather than filling a synced vault with recordings.

iPhone and iPad

• Shared transcripts now actually open on the device that didn't write them. Reading a transcript from the workspace folder went through without the folder's security scope and without asking iCloud to download the file — on the Mac that wrote the file both are free, on an iPad both fail silently, so a transcript showed its title and summary over an empty page. Reads now carry the scope, request the download, and say "fetching" or "not synced yet" instead of showing nothing.
• Shared transcripts can be played. The Mac has always archived the audio next to the markdown and named it in the frontmatter; the app now reads that and puts the same scrubber on a shared transcript that local recordings have.
• Transcripts render their structure — section headings, bold speaker names and bullets — instead of literal ## and **.
• Long transcripts no longer freeze the pane on open, and the library scan reads only each file's frontmatter instead of paging whole meetings into memory on every refresh.
• Search also matches a shared transcript's summary.

Mac

• A menu-bar icon of our own: the Transcripts mark — text lines over a waveform, matching the app icon — is the new default. It dims when auto-record is off, shows solid while watching, and while recording only its waveform pulses with the input level. The plain waveform and microphone styles remain in Settings, and the waveform style now dims when idle too.

1.0.0

The first release under this name.

Mac

• Menu-bar recorder with automatic call detection — captures your microphone and the other participants as separate tracks.
• On-device transcription and summaries. Apple Intelligence when present, a built-in summarizer when not.
• Live transcript written to a plain file during the call, so an assistant can read the meeting while it is still happening.
• Speaker attribution, correctable — reassigning a voice updates both the profile and the transcript.
• Automatic filing into your vault by keyword and on-device model.
• Import audio from anywhere — share a call recording, a Voice Memo or any audio file into Transcripts and it joins your library with its original timestamp.
• Sessions: group an evening's recordings under one name, with their own folder and an action that runs once when it ends. Start and end them from Shortcuts — with an optional label for the occasion and recording that begins straight away — so a weekly game can record itself.
• Defaults to iCloud Drive when you have it, so the Mac and the phone meet with nothing configured.

iPhone and iPad

• Records with the screen locked or another app in front, with a Live Activity on the Lock Screen and Dynamic Island.
• Live on-device transcription; refuses to transcribe rather than send audio to a server on a device that cannot do it locally.
• Hands recordings to the Mac through a shared folder — iCloud, OneDrive, Dropbox, or on-device.

Known limits

• Speaker names are a best-effort guess from voice and will occasionally be wrong. Settings ▸ Voices is how you correct them, and each correction retrains.
• The Mac app is unsandboxed by necessity: capturing other apps' audio and writing to a folder you choose are both outside the sandbox.
• It has been in daily use for months, but by very few people. Early Access is about widening that — the untested part is the variety of setups, not the app.