search

BUTSpeechFIT / VBx

Variational Bayes HMM over x-vectors diarization
251 stars 57 forks source link

what is the difference between plda_psi and plda_tr in line 93 and lines 97 and 98? #35

Closed xuanjihe closed 2 years ago

xuanjihe commented 3 years ago

image In vbhmm.py, what is the difference between plda_psi and plda_tr in line 93 and lines 97 and 98? I see that some lines in plda_tr are multiplied by -1.

fnlandini commented 2 years ago

Hi, Those lines refer to the transformation applied to the x-vectors explained in section 2.3. Equation 3 (solving the generalized eigenvalue problem) corresponds to line 96. The new value of plda_tr is the matrix E from the paper (eigenvectors) and plda_psi becomes the matrix with the eigenvalues. W and B are the within and between covariance matrices corresponding to the PLDA model which are passed to eigh for computing eigenvalues and vectors. If you look at line 130, you will see that the transformation to the x-vectors corresponds to the equation 2 in the paper.

If this answers your question, feel free to close this issue.

xuanjihe commented 2 years ago

Thank you very much!