People in your library, with Cadence Pro. Everyone in your speaker library now has a row under your recordings. Open one to see each recording they are in and how long they spoke, and the recordings they might be in, to confirm or turn down from one place. Also in this version: • A waveform of the recording, colored by speaker, takes the playback slider's place. Tap or drag on it to jump to that moment, or bring the slider back with the waveform button beside the controls. • On iPad, the waveform sits above the playback controls, with its button beside Brief.

Model downloads show how far they have got, in megabytes, rather than "Downloading…" alone. Also in this version: • Speech models move to a new folder the first time you open this version, and the ones you already have come along. • With the language on Automatic, Parakeet now records the language it heard, as Whisper does. • Clear a renamed speaker's name and save, and it gets its default name back, as on the Mac. • iPhone no longer offers Export for a recording where no speech was found. • Choosing a speaker count while speakers are off no longer marks a transcript out of date.

Recording groups, with Cadence Pro. Sort the library into groups such as Clients or Podcast: each group is its own section, with Ungrouped last, and you can collapse the ones you are not using. Drag a recording from one group to another, or long-press it and choose Group. New Group is beside Add. Also in this version: • Word highlighting stays on when you move quickly from a long transcript to a short one. • Deleting a Parakeet TDT v3 build while the other one downloads asks first, instead of stopping the download.

Transcribing now works offline from the first recording. Speaker labels need a one-time 22 MB download; without a connection, Cadence transcribes the recording without speakers and says why, instead of failing. Transcribe again once you are online to add them. Also in this version: • A failed transcription says why, rather than "Not transcribed yet". • A recording in which the model heard no words says "No speech found", naming the model. • Cancel on a transcript stops that recording only, not the whole queue. Cancel All is in the library when several are waiting. • Deleting a recording stops its transcription. • On iPad, a recording shared from the app beside Cadence appears straight away. • A shared recording is no longer lost when the library cannot save it, for example on a full disk. It is added the next time Cadence opens. • Deleting one Parakeet TDT v3 build asks first, since it removes the other build too. • Switching recordings no longer flashes the previous transcript. • A brief being written stays with its own transcript, and Retry generates for the right one. • On iPad, the model menu ticks the model on screen. • Settings credits the speech models, as their licences ask.

A new speech model in Cadence Pro: Parakeet Ultra. Parakeet Ultra is Parakeet TDT v3, further trained by moondream: the same 25 European languages with word timings, and fewer mistakes. In our tests it made about 15% fewer errors than v3 on recorded English calls, and on recordings several minutes long it never skipped a passage, where v3 occasionally drops a sentence. Choose it in Settings. It is a 632 MB download and sits beside v3 rather than replacing it, so your v3 and the transcripts made with it stay as they are. Also in this version: • New models are marked New in the model lists, for the release they arrive in. • Recordings left "Transcribing" or "Queued" when iOS closed Cadence in the background are put back the next time it opens. • A recording shared to Cadence is kept when it cannot be added yet, for example on a full disk, and added the next time Cadence opens. • A recording shared in can no longer lose its audio to the background cleanup. • On iPad, a recording opened beside its brief keeps its own name in the title bar, and deleting a recording closes it first.

A second speaker model, for recordings where people talk over each other. Nemotron 3 Diarization, in Cadence Pro, follows overlapping voices that the default model hears one at a time. It handles up to eight speakers and finds how many there are by itself. Choose it in Settings, under Speakers. It is a 201 MB download, and it does not keep voices, so the speakers it finds cannot join the speaker library: the transcript says so and suggests the other model. The free default model is unchanged. Also in this version: • A new app icon: the transcript over its player, in the speakers' colors.

Export transcripts from your iPhone. The Export menu, until now only on iPad and the Mac, is on iPhone too. Plain text, SRT and Copy as Plain Text are free, and WebVTT, Markdown, JSON and CSV are part of Cadence Pro. Save the file anywhere Files can reach, including iCloud Drive. Also in this version: • A speaker library, in Cadence Pro. Name a speaker and Cadence offers to recognize their voice in your other recordings. One tap confirms it, another turns it down, and nobody is named without that tap. Only a summary of each voice is kept, on your device, never the audio. Recordings transcribed before this update need transcribing again before they can be matched. • Shared recordings can start transcribing on their own. Turn on Transcribe shared recordings in Settings, under Sharing, and a recording you share from another app opens already transcribing whenever there is nothing to choose first. • Cadence has its own accent color, an amber-orange taken from the app icon. • Briefs on iPhone, in Cadence Pro: decisions, action items and open questions, in a sheet over the transcript, which keeps playing, so tapping a timestamp plays that moment. Briefs need Apple Intelligence, on iPhone 15 Pro or later. • Summarize one turn, in Cadence Pro: touch and hold a turn of 30 seconds or more and choose Summarize Turn for a few sentences on what that speaker said. It uses the same on-device model as briefs. • More room for the recording's name: the model, Transcribe Again and Settings move into the title's own menu, Brief moves beside Find in the floating bar, and Add Recordings sits beside search in the library. • The playback bar floats on glass, as it does on the Mac, and find lives in it: tap the magnifying glass to search the transcript. • While playing, the transcript keeps scrolling through a long turn instead of waiting for the next speaker. • Choose how strongly the spoken word is marked, under Transcript in Settings. In dark mode the default is now a step stronger. • A locked feature opens Cadence Pro right where you are, and long transcripts open and play more smoothly. • The speaker model downloads once more after this update (22 MB).

Read transcripts in a serif. A transcript is long-form prose, and Transcript in Settings now sets it in New York, the face Books and News use for reading at length. Speaker names and timecodes stay in San Francisco, so a turn is still easy to scan. There is a text size beside it, from Small to Extra Large, and it multiplies Larger Text rather than replacing it, so a reader who has already turned the whole system up gets both. Also in this version: • Stop or remove a queued transcription with a swipe, rather than red text on every row. A job that is already running can still be cancelled from its transcript screen. • Finished recordings no longer carry a green check in the library. Done is the resting state, so only recordings that need attention are marked: not transcribed yet, queued, running, failed, or missing their file. • The model menu uses the same glyph as the one on the Mac.

Cadence runs on iPad. The library and the transcript sit side by side, in any orientation, in Split View and in Stage Manager. Four things the iPhone leaves out come with it: drag and drop from Files, Voice Memos or Mail; export to every format; the brief, in an inspector beside the transcript; and two models compared side by side, with the differences marked. Also in this version: • Share audio or video to Cadence from the share sheet: a memo from Voice Memos, an attachment in Mail or Messages, or a file in Files. • Each recording picks its own model, language and speaker count, in a Transcribe sheet that lists every model with its download size. • Keyboard shortcuts, with a hardware keyboard: ⌘O to add recordings, ⌘F to find in a transcript, Space to play or pause, ⌥← and ⌥→ to skip 15 seconds. • Language lists are sorted by name rather than by language code, so German no longer comes before English. • Exports are saved with the right file extension, so a subtitle file opens in a video player without being renamed. Briefs need Apple Intelligence.