Describe the bug

When I evaluate this model on VoxCeleb-O, I get an EER around 8.9%, rather than 3.2%. I do not see the previously reported issue about all scores being high: in fact, the EER threshold is 0.94. Remarkably, I do replicate the reported EER on spkrec-ecapa-voxceleb and spkrec-resnet-voxceleb, with the same code.

Expected behaviour

Performance Match

To Reproduce

No response

Environment Details

No response

Relevant Log Output

Additional Context

No response