"VoxCeleb Xvector System 1a" pretrained model - PLDA scoring has Less Variability between Enrolled and Test Speakers

[Immu Ali]

Hi dan ,

Previously i used "SRE16 Xvector Model 1a" pretrained model and the PLDA scoring results are Good i.e high variation between the test and Enrolled Speakers 
 

Currently I'm using "VoxCeleb Xvector System 1a" pre-trained model and Found out that , there's less variability between Enrolled and Test Speaker , So , I checked the Recipe "run.sh" Script and Found out that in PLDA stage i.e, "stage-11" we are passing test speaker xvector two times  during PLDA scoring whereas in SRE16\V2 Recipe "run.sh" we are giving 1.test speaker xvector and 2.Enroll Speaker Xvector .

So , How Exactly PLDA Classifier Works in VoxCeleb\v2 Recipe by passing test Speaker xvectors 2 times ?   

Dummy Question : Why we are not giving Enroll Speaker Xvector in Voxceleb\v2 Recipe? 

Thanks in advance.

[David Snyder]

It's nothing to worry about. It's just due to the way the evaluation is defined in the trials file.

In SRE16, it's something along these lines:

enroll-spk1 test-utt1 target

enroll-spk1 test-utt2 nontarget

enroll-spk2 test-utt1 nontarget

...

In Voxceleb it's something like this:

test-utt1 test-utt2 nontarget

testutt1 test-utt3 target

test-utt2 test-utt3 nontarget

...

So, the Voxceleb trials use the same list of utterances for both the first and second columns of the trials file, which is while the test utterances are passed into the PLDA scoring script twice. However, in SRE16, the evaluation is split into an enroll and test portion.