SuperSubtitler v0.3.1

SuperSubtitler 0.3.1 — speaker recognition, measured and tuned

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

🗨️ Speaker changes: four times fewer mistakes

• “Mark speaker changes” (-speakers) is now measured: a new test runs it on 66 minutes of hand-labelled debates and news with 1 to 11 speakers (VoxConverse) and scores every result.
• Six speaker models at many settings were compared. The winner, 3D-Speaker ERes2Net at speaker_threshold 0.95, gets it wrong 10.8 % of the time — 0.3.0's model at 0.5 got 44 %, mostly by splitting one person into several (which meant “- ” lines where only one person spoke).
• The new model (26 MB) is downloaded once, pinned and checked like everything else, from the project's own mirror first.
• If your settings still hold 0.3.0's default (speaker_threshold=0.50), it is moved to the new default automatically; a value you chose yourself stays. The Settings page shows the new recommended range (0.85–1.05).
• CI now checks the error rate on every change to the speaker recognition, so it cannot quietly get worse.

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