When you trained the model on voxceleb, did you balanced the number of files per speaker ?

[amitli1]

• The ECAPA-TDNN model trained on "development part of the VoxCeleb2".
• The datasets contains 5994 speakers.
• Some speakers have many more files than other speakers (imbalanced classes).
• Did you handled this imbalanced ?

[TaoRuijie]

No, this imbalance is in a proper range.