One overlap aware tweak that genuinely moved the needle for me was increasing the overlap penalty during VBx resegmentation while tightening the prior on speaker turn duration. In multilingual meetings, especially ones mixing English and Hindi or Spanish, the default settings tended to over assign speech to a single dominant speaker whenever short overlaps occurred. That inflated speaker confusion even when boundaries were mostly right. The concrete change was raising the overlap penalty from 0.3 to 0.6 in the VBx stage and slightly lowering the minimum speaker duration prior so short interjections were not smoothed away. This helped the model tolerate brief overlaps without collapsing them into one speaker segment. I validated this on a 42 minute internal meeting recording with four speakers switching languages mid sentence. Before the tweak, DER was 14.8 percent. After the adjustment, DER dropped to 11.9 percent. The biggest gain came from speaker confusion, which fell by almost three points, while false alarm and miss rates stayed roughly flat. One specific clip stood out. Around minute 18, two speakers overlapped for about two seconds during a language switch. Previously, VBx assigned the entire region to one speaker. With the higher overlap penalty, the segment was split correctly and attributed to both speakers with much cleaner boundaries. The lesson I took from this is that overlap handling matters more than embedding quality once you are in real meeting audio. Multilingual speech naturally increases short overlaps and backchannels. Penalizing overconfident merging while allowing brief turns made the diarization more human like and measurably better, without retraining embeddings or changing clustering entirely.