search

TaoRuijie / ECAPA-TDNN

Unofficial reimplementation of ECAPA-TDNN for speaker recognition (EER=0.86 for Vox1_O when train only in Vox2)
MIT License
581 stars 111 forks source link

MinDCF result #80

Closed RMobina closed 4 months ago

RMobina commented 4 months ago

In the paper, the minDCF results are reported with P_target=0.01, but in the source code is set to 0.05.

minDCF, _ = ComputeMinDcf(fnrs, fprs, thresholds, 0.05, 1, 1) I would appreciate it if you could clarify whether the results are based on a P_target of 0.01 or 0.05.

also, I have developed different extended versions of ECAPA-TDNN model with p_target = 0.05, and I want to evaluate how it performs with a p_target = 0.01, should I simply adjust the p_target = 0.01 and evaluate my pretrained version with python trainECAPAModel.py --eval --initial_model exps/pretrain.model or if is necessary the model trains again with p_target = 0.01?

Thank you.

TaoRuijie commented 4 months ago

p_target is only related to minDCF, I use EER as the main metric so I did not care it too much.

Yes, for the trained model, just adjust P_target during evaluation is ok.

RMobina commented 4 months ago

Thank you for your clarification.