search

nikvaessen / w2v2-speaker-few-samples

Research code for the paper "Training speaker recognition systems with limited data" at https://arxiv.org/abs/2203.14688
MIT License
11 stars 2 forks source link

EER of Ecapa on vox1-o #2

Open WhXmURandom opened 1 year ago

WhXmURandom commented 1 year ago

Hi,I have a question and I hope you can answer it.Generally speaking, the EER of ECAPA trained on Vox2 and tested on Vox1-o is around 1%. Why is your result 2.91%?

nikvaessen commented 1 year ago

We do not do data augmentation nor any post-processing of the speaker embedding.

WhXmURandom commented 1 year ago

What specifically does "post-processing of the speaker embedding" refer to? Are you implying that you did not perform any score normalization or calibration when calculating the EER?

nikvaessen commented 1 year ago

Are you implying that you did not perform any score normalization or calibration when calculating the EER?

yes, I did not do any normalization/calibration.

WhXmURandom commented 1 year ago

Thank you very much for your answer！