pre-trained PLDA model

[賴禹邵]

Hi, I've downloaded the pre-trained plda model and used it to implement a x-vector-based speaker verification task. The model is downloaded from openslr site:
https://kaldi-asr.org/models/m7 (VoxCeleb Xvector System 1a)

It seemes that this model has only three parameters, including 2 vectors and 1 matrix, while the plda I'm familiar with should be a 4-parameter set:
m: global mean (vector)
W: Whitening transformation (matrix)
V: Basis of eigenvoice (matrix)
Σ: Covariance of noise of plda (matrix)

Can anyone tell me how this model file corresponds to these 4 parameters?
Thanks.

[David Snyder]

You can use ivector-copy-plda with --binary=false to copy the model into a plain text file. See https://github.com/kaldi-asr/kaldi/blob/master/src/ivectorbin/ivector-copy-plda.cc.

Take a look at the Write method for the PLDA object: https://github.com/kaldi-asr/kaldi/blob/master/src/ivector/plda.cc#L26 .

Each of those objects (matrices and vectors) are described in comments in the PLDA header file: https://github.com/kaldi-asr/kaldi/blob/master/src/ivector/plda.h#L148

You may need to read "Probabilistic Linear Discriminant Analysis" by Sergey Ioffe, to understand some of the code. This is what Dan used as a reference when writing it. You might be more familiar with Prince's notation.