Odyssey 2012 - The Speaker and Language Recognition Workshop

June 25-28, 2012

Speaker Vectors from Subspace Gaussian Mixture Model as Complementary Features for Language Identification

Oldřich Plchot (1), Martin Karafiát (1), Niko Brümmer (2), Ondřej Glembek (1), Pavel Matějka (1), Edward de Villiers (2), Jan "Honza" Černocký (1)

(1) Brno University of Technology, Speech@FIT and IT4I Center of Excellence, Czech Republic
(2) AGNITIO, South Africa

In this paper, we explore new high-level features for language identification. The recently introduced Subspace Gaussian Mixture Models (SGMM) provide an elegant and efficient way for GMM acoustic modelling, with mean supervectors represented in a low-dimensional representative subspace. SGMMs also provide an efficient way of speaker adaptation by means of lowdimensional vectors. In our framework, these vectors are used as features for language identification. They are compared with our acoustic iVector system, which architecture is currently considered state-of-the-art for Language Identification and Speaker Verification. The results of both systems and their fusion are reported on the NIST LRE2009 dataset.

Full Paper

Bibliographic reference.  Plchot, Oldřich / Karafiát, Martin / Brümmer, Niko / Glembek, Ondřej / Matějka, Pavel / Villiers, Edward de / Černocký, Jan "Honza" (2012): "Speaker vectors from subspace Gaussian mixture model as complementary features for language identification", In Odyssey-2012, 330-333.