questions about ECAPA-TDNN model training

[lawlict]

questions about ECAPA-TDNN model training

[lawlict]

Hi @forwiat. The outstanding performance of ECAPA-TDNN doesn't fully rely on the model structure. In fact, there are lots of details that help improve the performance. For example:

• 6 kinds of data augmentation
• a cyclical learning rate
• adaptive s-norm

I recommend that you use this model based on voxceleb_trainer(https://github.com/clovaai/voxceleb_trainer). Still, I only achieved a EER of 1.2% on the original voxceleb1_test set. Therefore I guess that some details are still missing.

[lawlict]

Hi, @TaoRuijie. Yes, your understanding is correct.

[forwiat]

OK, I will carefully try it again!
Thanks for your quick reply and answer!

[TaoRuijie]

Hi, thanks for your sharing!

So you are using the same code in voxceleb_trainer, instead of changing the model into this ecapa-tdnn right? And you can get 1.2 EER in vox1 test set. Is my understanding correct?

[seacj]

Hi, I also train the ECAPA-TDNN model (C=1024) with voxceleb_trainer but get 1.19 EER . After using as-norm, I get 1.08 EER.
Do you have any suggestion to get 0.87 EER in the paper?

[lawlict]

Hi @forwiat. The outstanding performance of ECAPA-TDNN doesn't fully rely on the model structure. In fact, there are lots of details that help improve the performance. For example:

• 6 kinds of data augmentation
• a cyclical learning rate
• adaptive s-norm

I recommend that you use this model based on voxceleb_trainer(https://github.com/clovaai/voxceleb_trainer). Still, I only achieved a EER of 1.2% on the original voxceleb1_test set. Therefore I guess that some details are still missing.

Hi, @seacj. As you see, I only achieve a EER of 1.2% on Voxceleb-1-O.trials without as-norm. I guess that the original paper uses Voxceleb1-O-clean.trials, which is slightly easier.