Release notes

Interviews file into projects, and Kayit arrived on iPad.

• Kayit runs on iPad. The same app as the iPhone one and the same purchase as the Mac one. On an iPad it opens as three panes, with your interviews down one side, the transcript in the middle and the on-device insights beside it, which is the one part of the app a phone does not have.
• Interviews can be filed into projects. A project is a folder in your library, so your filing is the same in Kayit, in the Finder and in a backup, and nothing moves unless you ask. Choose New Project from the File menu, drag an interview onto a project in the sidebar, or use Move to Project from a row. Archive the project when the story is filed.
• A project carries one roster of speaker names. Naming a speaker offers the names you have already used in that project's other interviews, so a source you have interviewed six times is spelled the same way six times.
• Every quote in a project sits in one list. Open a project's quotes to see every passage you have saved across its interviews, and export them together.
• A project keeps its own notes. The running list of what you still need lives in the project as a plain Markdown file, so it opens in the Finder or in your notes app as well as in the app.
• A whole project exports the way a single interview does. Audio and transcripts travel together, and the project opens on another Mac.
• The sidebar filters by name. Type in the field below the list to find an interview wherever it is filed, including inside an archived project.
• Renaming an interview takes everything saved beside it along. Choose Rename Interview from the row menu in Recent or All Interviews, or from the Transcribe window, and the recording, the transcript, your saved quotes and the insights all move with the new name.
• Correcting a typo no longer discards the insights. Editing a single character used to throw away the summary and every flagged passage for that interview. They survive the edit now, and stay on the right words.
• The insights can be copied out. Copy or share the whole summary as Markdown, or use the context menu on a single point or outline entry to take just that one. A Newsworthy passage copies with its speaker and timestamp, in the same shape a saved quote copies in.
• You name what you export. The export sheet lets you name the text file, the zip or a single audio clip yourself, instead of taking the name that was worked out for you.
• Insights runs the better model on more Macs. The most capable on-device model needed 24 GB of memory and now needs 16 GB, so on a 16 GB Mac the panel finishes in about two-thirds the time and chooses better. A Mac with 8 GB now runs the mid-sized model, which used to need 14 GB.
• More passages get flagged, and fewer of them repeat. The pass that finds striking lines was throwing away about half of what it found, and the summary no longer repeats quotes the Newsworthy list already carries.
• Hiding a speaker hides everything they said. Filtering a speaker out of the Newsworthy list used to leave their own striking lines standing underneath it.
• The speaker's name sits above the paragraph, and appears only where the speaker changes. A run of paragraphs from one person reads as one turn. The waveform is legible in dark mode again.

Transcribing became free, and Kayit arrived on iPhone.

• Transcribing is free. Transcription, correcting the text and exporting it as plain text no longer need a subscription, and there is no time limit on them. Kayit Pro adds what is built around the words: speaker separation, My Voice, saved quotes and their audio, the on-device insights, timestamps in an export, and export as Markdown, CSV, SubRip subtitles or a file carrying the audio. Nothing changes for anyone already subscribed.
• There is an iPhone app. Transcribe on the phone with the same three engines, the same speaker separation and the same saved quotes. One Kayit Pro subscription covers both apps, because it is one purchase on one App Store record.
• My Voice matches your voiceprint directly. Your saved samples used to be added to the front of the recording so the app could work out which voice was yours. It matches the voiceprint itself now: transcribing is faster, the speaker count describes the interview and nothing else, the transcript is the same whether My Voice is on or off, and your name survives a re-run. Label My Voice applies it to a transcript you already have.
• Saved quotes hold their place. A quote is anchored to its moment in the recording, so correcting a typo, splitting or merging paragraphs and re-running speaker separation all leave it sitting on the words it was taken from. If a re-run draws a new paragraph break through a quote, the paragraphs are rejoined under the speaker the quote already carried.
• You can jump to a timestamp. Click the elapsed time in the player and type one.
• Speaker names match however you type them. Naming a paragraph "daniel" where Daniel already speaks assigns it to Daniel instead of creating a second speaker one capital letter away, and renaming a speaker onto a name already in use warns that it would fold two people into one. Naming every unlabeled paragraph at once can be undone.
• Merging paragraphs asks who is speaking. In edit mode, pressing Delete at the start of a paragraph that carries a different speaker from the one above asks which name survives, instead of keeping the one above without asking.
• A long step says what it is doing. If work goes quiet for fifteen seconds, the Process log names what it is still on and how long it has been there, and loading a model quotes what that took last time on this Mac instead of promising it will be quick.
• Adjusting what gets flagged is one decision. The Insights lens has a single bar, set to flag a claim at weak, medium or strong, plus an on/off switch for each category, in place of a bar per category. A lens tuned in an earlier version carries across rather than resetting.
• The flagging note has its own switch. Settings, On-Device AI now carries a second checkbox beside the summary one. Turn it off and a flagged passage keeps the speaker's own words and where they fall in the recording, with nothing the model wrote left in the panel. It also clears the note from interviews already analyzed, and an organization can enforce either switch with a managed profile.
• Setup no longer asks which speaker separation method to use. Pyannote measured faster and more accurate on every interview tested, so it is the one that runs, and FluidAudio stays in the re-run panel for a second opinion. Settings, Offline lists both as separate downloads so you can take them both before losing your connection.
• Download sizes are the real ones. Engine and speaker separation cards quote what will actually be fetched rather than a conservative estimate, and say whether the model is already on this Mac.
• Start takes hold the moment you press it. It used to sit inert while a model downloaded, so it got pressed again, and each press started another run over the first.

Getting audio in, and getting the speakers right.

• Audio can be sent straight in from another app. Kayit appears in the macOS Share menu and in the Finder's Open With, so an interview goes across from Voice Memos, Photos, Music or a folder. Apps that hand over a file lazily used to be ignored without saying so; their files are accepted now.
• Video files import. Drop in an mp4, mov or m4v and the audio track is pulled out for transcription.
• Speaker separation can be re-run. Correct the speaker count and run it again on a transcript you already have. Re-transcribing with a different engine keeps your saved quotes too.
• Save a sample of your own voice. Record yourself once and your turns are labeled with your name in later interviews rather than as Speaker 1.
• Newsworthy passages are flagged, and stay flagged. The panel marks claims worth a second look and writes them to disk, so they are still there the next time the interview is opened. Entity matching got better at first names, full names, and telling a place from a person.
• The insights moved from an inspector into a side panel. It sizes itself properly when it is opened and reopened, which the inspector never quite did.
• Models come from Apple rather than Hugging Face. Transcription models are delivered as Background Assets packs over Apple's content delivery network, with a real percentage while they fetch, falling back to Hugging Face if that is unavailable.
• Setup walks through the downloads up front. A rebuilt onboarding flow picks an engine and takes what it needs ahead of time, so the first interview is not waiting on a model.

The first release, on the Mac App Store.

• On-device transcription for interviews. Three engines and a hundred languages, running on your own Mac, with playback synced to the words and an editor where correcting a word keeps its timing.
• Speaker separation. The transcript splits each time the speaker changes, and renaming a speaker applies across the whole interview.
• Quotes. Select a passage to save it with its speaker and its place in the recording, then edit or export a batch of them together.
• Insights that run on the same Mac. An optional summary, an outline and verbatim key quotes, produced by a local model, with a switch that turns all of it off.
• Export and import. Export a transcript in several formats, drag audio in, and reopen an interview instantly from a cached waveform.