Publication Details
Full-covariance UBM and Heavy-tailed PLDA in I-Vector Speaker Verification
Glembek Ondřej, Ing., Ph.D.
Castaldo Fabio
Alam Jahangir
Plchot Oldřich, Ing., Ph.D. (DCGM)
Kenny Patrick
Burget Lukáš, doc. Ing., Ph.D. (DCGM)
Černocký Jan, prof. Dr. Ing. (DCGM)
GMM, speaker recognition, PLDA, heavytailed PLDA, full-covariance UBM, i-vectors
The work we presented aims at the best performance of the single stand alone system. We have presented full-covariance UBM and i-vector extraction with different kind of modeling. Our analysis shows that for the best performance it is necessary to have fullcovariance i-vector without any approximation.
In this paper, we describe recent progress in i-vector based speaker verification. The use of universal background models (UBM) with full-covariance matrices is suggested and thoroughly experimentally tested. The i-vectors are scored using a simple cosine distance and advanced techniques such as Probabilistic Linear Discriminant Analysis (PLDA) and heavy-tailed variant of PLDA (PLDA-HT). Finally, we investigate into dimensionality reduction of i-vectors before entering the PLDA-HT modeling. The results are very competitive: on NIST 2010 SRE task, the results of a single full-covariance LDA-PLDA-HT system approach those of complex fused system.
@inproceedings{BUT76387,
author="Pavel {Matějka} and Ondřej {Glembek} and Fabio {Castaldo} and Jahangir {Alam} and Oldřich {Plchot} and Patrick {Kenny} and Lukáš {Burget} and Jan {Černocký}",
title="Full-covariance UBM and Heavy-tailed PLDA in I-Vector Speaker Verification",
booktitle="Proceedings of the 2011 IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP 2011",
year="2011",
pages="4828--4831",
publisher="IEEE Signal Processing Society",
address="Praha",
doi="10.1109/ICASSP.2011.5947436",
isbn="978-1-4577-0537-3",
url="https://www.fit.vut.cz/research/publication/9657/"
}