Skip to content

Dictate 2.6.0

Choose a tag to compare

@Budanovvv Budanovvv released this 19 Aug 21:38
· 117 commits to main since this release

Meetings you can read, search, and get out of the way of.

One person stays one person. The model that finds speaker changes can only
see ten seconds at a time, so every longer window was cut into pieces — and the
last piece, padded out with silence, was embedded as a stranger. On a real
forty-minute call that split one voice into two near-equal "speakers" across
70% of windows, which is why a monologue arrived in the transcript alternating
between two names. Windows are now capped at what the model actually sees.
Measured on the same recording: 6% instead of 70%, and the talking split out
85/14/1 — which is what the meeting actually sounded like.

And when a sentence does get torn, the words decide who owns it. A line
that stops mid-phrase followed by one that opens in lower case is one person
being interrupted by the clock, not two people talking, and it is now filed
that way.

A recording can get out of the way. Fold the transcript into a pill — level
meter, clock, stop — with the yellow button or by closing the window; close the
pill and the menu bar is all that is left, where the mark breathes for as long
as the meeting runs. The menu brings the transcript back at any point.

Copy anything. Every turn has a copy button and a right-click menu — the
words alone, or the line with speaker and time — and ⌘C works in the window
without stealing focus from your call. Selecting text no longer fights the
transcript: it stops auto-scrolling the moment you scroll up or start a
selection, and a button brings you back to the newest line.

A transcript that reads like a conversation. Recognition cuts audio into
windows, and every window used to become its own line, so sentences broke
mid-phrase and hums stood alone as if they were turns. They are now joined back
into paragraphs — 1204 lines became 693 in a typical archive — while the file
on disk keeps every word exactly as it was recorded.

Contents and summaries. Each meeting gets a one-line summary of what it was
about and a table of contents with timestamps, written into the Markdown file
itself, so the folder is useful without the app.

Search that understands meaning. Ask for "legal" and you also get the
meeting where nobody said the word. Type a question in any language; it is
matched against what the meetings were about, and clicking a result opens the
transcript at that moment.

Titles and summaries are written on your Mac by a model you can choose to
download
— 2.5 GB, one time, Apple Silicon and Intel alike. It replaces the
system model that only exists on macOS 26 and refuses to summarize anything it
deems sensitive, which in practice meant medical and legal conversations came
back blank. Nothing leaves the machine either way.

The library opens instantly. It used to read every transcript before
showing anything, on the thread that draws the app — and in an iCloud-synced
Documents folder a read can mean a download. Opening it as a call started
froze everything for sixteen seconds and cost the first twenty seconds of the
meeting.

The consent reminder appears once, before the first meeting you record,
instead of before every one.

Full Changelog: v2.4.0...v2.6.0