SuperSubtitler v0.3.2

SuperSubtitler 0.3.2 — speaker changes: more quick replies, much faster

Download both files into one folder: SuperSubtitlerGUI.exe (the window) and SuperSubtitler.exe (the engine).

🗨️ “Mark speaker changes” works differently now

• Testing showed the weak spot: two people speaking quickly one after the other often went unnoticed — one subtitle, two speakers, no “- ” lines.
• The reason: until now the voices were told apart across the whole film. That needs plenty of speech per person, so a one-second reply was simply counted to whoever spoke before.
• Subtitles never show who speaks, only where the voice changes. So that is all SuperSubtitler looks for now: the pyannote segmentation model, trained exactly for this, judges every second about nine times, and a change counts where enough of those looks agree.
• Measured on hand-labelled recordings (66 minutes of debates and news, 1–11 speakers):

	changes found	quick replies found	found changes that are real
0.3.1	67 %	46 %	73 %
0.3.2	74 %	59 %	73 %

• About 8 times faster: a 2-hour film needs roughly 2 minutes instead of 10–15.
• Smaller: only a 6 MB model; sherpa-onnx and the speaker embedding are no longer needed.
• New setting How sure a speaker change must be (speaker_change_threshold, 0.15 measured best). The old speaker settings of 0.3.0/0.3.1 are replaced automatically.
• Honest limit: quick replies are the hardest case for every system; about 4 in 10 of them still go unnoticed.

The exes are not code-signed; Windows SmartScreen may warn — "More info" → "Run anyway". Built by GitHub Actions.