CaptureMann 0.2.3

0.2.3 is about speaker detection, plus the Intel Mac recording crash and the fixes found while testing 0.2.2.

Fixed

• Intel Macs: starting a screen recording no longer aborts the app (GitHub issue 16, MacBookPro16,1 with an Intel UHD 630). AVFoundation rejected the encoder settings by throwing an Objective-C exception, which unwound into Rust and killed the process, leaving the "Currently Sharing" indicator up. The encoder set-up (and every frame append) now runs inside an exception catch. A rejected plan is tried again without the optional compression properties, then (for HEVC up to 4096 px wide) as H.264; the manifest records the codec the file really holds. When nothing works the recording says the Mac has no usable hardware encoder instead of crashing, and the reason AVFoundation gave is written to the system log. Not yet confirmed on the reporter's Mac.
• A short Vietnamese recording came out as Indonesian ("Hai pinta Recalling screen dan bantuan Jemaah..."). Whisper's language detector listens to the first 30 seconds and, on an 8-second clip, rated Indonesian 55 % and Vietnamese 24 %. A new setting, Languages you speak (AI page, ai.languages, for example vi, en), makes the detector choose only among those; the same clip then reads "kiểm tra Recalling screen". Also language: "vi,en" on the transcribe tool and capturemann ai transcribe --language vi,en. Detection now skips the silence before the first sound. Until the setting is filled in nothing changes: Whisper still picks among all its languages. To repair an old transcript, set the languages and Re-generate it.
• Reveal in Finder on a recording or an audio capture failed with "not a screenshot".
• Cancel on a report left "Writing the report 0%" on screen for ever; the card now clears once the job has stopped.

Changed

• Speaker detection uses a better voice model. The English-only CAM++ model is replaced by the 3D-Speaker CAM++ Chinese-English model (Apache-2.0, 28 MB), and the clustering threshold for an unknown number of people is 1.0 instead of 1.2. On 6 public English conversation clips held back while choosing, the diarization error rate fell from 54 % to 17 % (51 % for calling everyone one speaker) and the speaker count was right in 5 of 6 clips, against 0 of 6. English only: Vietnamese is not measured yet. Details and caveats: verification/benchmarks/speaker-detection-2026-10-29.md. Speakers found by 0.2.2 stay as they are until you run Find speakers again.
• Finding speakers shows real progress instead of 0 % until it ends. Cancel still takes effect when the run ends (the speaker library ignores a request to stop).

Added

• Languages you speak (AI page, setting ai.languages): see the transcript fix above. The transcribe tool and capturemann ai transcribe take a list such as vi,en.
• A benchmark folder, verification/benchmarks/, with the speaker-detection benchmark (setup, public data, results, what it does not show) and the tools behind it: word and diarization error rates (cm-ai measure), the baseline runner (baseline), a draft-reference maker for your own recordings, tests/tools/make_reference_clips.py for PriMock57 and AMI clips, and lang_probe, which prints Whisper's language probabilities. None of it ships in the app.

Known limitations

• Speaker accuracy is measured on 14 English conversation clips only. Vietnamese and other languages are not measured, so the accuracy criteria of speaker labels stay open. The speaker count is still a suggestion: 3 of 8 tuning clips and 1 of 6 held-out clips got it wrong, usually one speaker too many.
• Finding speakers on a very long recording uses a lot of memory (about 4.4 GB for 88 minutes) and Cancel takes effect only when the run ends: the speaker library ignores a request to stop. Running it in windows would fix both and is not done.
• The Intel recording fix is untested on a real Intel Mac. If every retry is rejected, recording still fails (with an error instead of a crash); the reason is in the system log.
• The language list only helps once it is filled in; with the field empty a short clip can still be heard as a related language.
• The known limitations of 0.2.2 still apply (no transcript word error rate for Vietnamese, the second reading only on loops, longer runs for long recordings).