Performance of the new Segmentation model

[omarsayed7]

Hello all,
I measured the DER on custom dataset of the new Segmentation model with WeSpeaker_ResNet34_LM variant and got 19.4%, while the old Segmentation model from v2.1.1 with WeSpeaker_ResNet34_LM variant and got 16.9%.
Could you please explain why the new Segmentation model gets bad results compared to the old one?

Regards,
Omar

[hbredin]

Marking this as cannot_reproduce until you provide some example files (and corresponding manual annotations) where this happens.