Speaker labels used to arrive after the note did. Now they arrive with it.
The wait is gone
Finding out who was talking needs the audio and nothing else — it listens to voices and never reads the words. It was still being done last, though: a ten-minute interview transcribed for three minutes, appeared on screen unlabelled, and only then started a two-minute pass whose answer had been available from the first second.
It now runs alongside the transcription, so the note's first appearance carries its labels. Nothing on screen says it is happening, because there is nothing to wait for.
It does not slow the transcription down. The two use different parts of the chip on purpose. Measured on the Mac this was built on — three interleaved rounds, twelve seconds of speech with a ten-minute pass running beside it:
alone 1506 / 1531 / 1497 ms
listening 1586 / 1569 / 1576 ms
66 milliseconds, about 4%. Your own numbers will differ; scripts/bench.sh is in the repository so you can take them.
Naming a speaker
It looked like it did nothing. Typing a name over "Speaker 1" rewrote every turn immediately — the note on disk was correct — but the words in front of you did not change until you left the note and came back. A name with a space in it looked worst of all, as though the space had been rejected. It never was. That was the same invisible update wearing a second hat.
Your place in the note survives it: the scroll position, the paragraph you were reading and anything playing are all left alone.
And it offered you the label you were trying to get rid of. The suggested names are read out of what was actually said aloud, by a model reading the transcript — and a transcript whose every paragraph opens with the word "Speaker" leads it straight to "Speaker 1". It no longer offers a name it invented from the labels.
Everything that moves
The app's motion is bound to the design system rather than to whatever each screen was written with. That was nine different entrances — rises of 8, 10 and 14 pixels, durations from 0.28 to 0.42 seconds, five different staggers — all on one curve, because each was tuned in the file it lived in. They are one gesture now, and a panel arriving no longer shares an ease with a progress bar filling.
- Long lists finish arriving in the same time as short ones instead of getting slower as your library grows.
- Counting numbers step at a rate you can read rather than sixty times a second.
- The reading view's text resize is unremarkable on purpose now, which is what keeps your eye on the line you were on.
Also
- The SPEAKERS button changes what it says while it works, and it was pushing DELETE sideways every time — with the pointer already on its way there. It is laid out at its widest now and nothing moves.
- Renaming a speaker on a dictation said "Nobody in this meeting is called that". Since 1.2.0 those labels are on ordinary recordings too, and the wording had not caught up.
Requires an Apple Silicon Mac. Everything runs locally; nothing is uploaded.