Transcripts checked against the audio, Polish on this Mac and a reason for every change

Transcripts can be checked against the audio before they are pasted — turn on Settings - Performance - Audio verification preview to try it. After you stop, the places a transcript usually goes wrong — where one sentence meets the next, and words the model was unsure of — are listened to again, and the verified text is what gets delivered. Final transcript sets how much is rechecked: Fast finishes the remaining audio, Balanced checks short recordings and uncertain passages, Maximum rechecks the recording with your selected model. Compare keeps delivering the way it does today, then rechecks in the background and shows any difference in Review. It never pastes a second time and never overwrites an edit you made. The preview starts Off and stays opt-in while it is measured on real voices. The Final transcript row says when your choice is not in effect.

Polish: grammar tidied by a model on this Mac — a new Polish mode on Dictate. Import an instruction model file (GGUF) under Settings - Speech - Local Polish; nothing is downloaded or switched on for you. Message and Prompt use the same model once one is selected. Names from your vocabulary, numbers, versions, identifiers, file paths, links, email addresses and quoted words are checked before anything is pasted. If Polish still changes one after a retry, or takes longer than four seconds, the Clean transcript is delivered instead. The overlay shows Refining while it works. Press Tab to paste the Clean transcript straight away; the words shown while it writes are never pasted.

Styles by application — give an app its own mode and tone, such as a short Message for chat and Polish for email. The first matching row wins, and an empty row sets the default. Edits you make to a Polish result are kept as style examples, never as automatic word swaps, and Forget style examples clears them.

See what changed in a transcript, and why — Review under a transcript on Dictate, and on every clip in Review, opens the verified transcript and each change beside what made it: your correction memory, a context pack, Clean or Polish. Keep, Revert or Always use each change. The same correction repeated in one clip is grouped, with Keep all and Revert all. When the audio check could not settle a passage, the wording that was kept and the alternative both show, rather than one being picked for you.

Correction intelligence in Review - Analysis — built-in analysis reads what you corrected and shows what is improving and which mistakes keep coming back, all on this device. Its proposals — add vocabulary or a project term, turn off a rule that misfires, adjust the language plan — do nothing until you select them and choose Apply selected, and an applied set can be undone. An optional deeper pass can run on the local Polish model, or on a coding CLI you already have installed. You see exactly what would be shared before anything runs — see Privacy. Insights lists the suggestions you accepted and whether each is Helping, Unproven, Redundant or Needs review. Settings - Agents is now Correction intelligence: pick the provider for the deeper pass and how long a run may take.

Run in background — a button in the title bar and in Settings - General hides the window while the shortcut keeps working from any app. The first time, it explains how the window comes back: the menu bar or tray icon, or the Dock icon on macOS.

Measured model advice — import two evaluation reports made on the same checked recordings under Settings - Performance to see which model and preset made fewer errors, and what it costs in finishing time. It never changes your model.

Neural speech boundaries (experimental) — Settings - Performance can use an imported speech-boundary model to find where sentences start and end, with the built-in detection kept as the fallback.

More than one context pack can feed dictation — mark several projects as in use and their terms are ranked together, instead of only the first pack counting. Transcribing a clip again uses the same set.

Performance shows where the time went after you stop — the timing panel has separate Clean and Polish stages, and each recent dictation lists which check ran, how many passages were retried and how many fell back.

A recording cut short by a crash is explained on the next launch — the status bar says the unsaved audio and text were discarded. Nothing is resumed or pasted automatically.

Less background work while the window is hidden — microphone previews close and the overlay and captions clocks stop ticking until the window is shown again. On Windows, hidden windows also stop drawing.