Reproducing WavLM results on speaker verification

[stevehuang52]

Hello,

I'm trying to reproduce the speaker verification results in the WavLM paper using the ECAPA-TDNN baseline, but cannot get close to the scores in the paper. Could you please provide more details (e.g., data processing/augmentation, optim params, scheduler, epochs, etc) on training the base ECAPA-TDNN model as well as the one with WavLM features so that I can try again?

Thanks,
Steve