16th Annual Conference of the International Speech Communication Association

Dresden, Germany
September 6-10, 2015

Phonemes Frequency Based PLLR Dimensionality Reduction for Language Recognition

Saad Irtza (1), Vidhyasaharan Sethu (1), Phu Ngoc Le (1), Eliathamby Ambikairajah (1), Haizhou Li (2)

(1) University of New South Wales, Australia
(2) A*STAR, Singapore

This paper presents a new approach to reduce the dimensionality of Phone Log likelihood Ratio (PLLR) features, which have been shown to be effective for language recognition, by removing the likelihoods corresponding to less frequent phonemes. In this work, phoneme frequencies are estimated using a suitable phoneme recogniser. Following this, an i-vector framework is used to represent the total variability in the reduced dimensional PLLR feature space. This paper also proposes the use of Gaussian probabilistic linear discriminant analysis (GPLDA) as a backend for Language Recognition Evaluation (LRE) tasks. The suitability of both, the proposed dimensionality reductions technique and the GPLDA back-end has been evaluated on NIST 2007 and 2011 LRE tasks. The results show that the novel dimensionality reduction method outperforms PCA based dimensionality reduction by 7%. Further the results also show that GPLDA outperform generatively trained Gaussian back-ends, which have previously been used in conjunction with PLLR feature, by 14.6%.

Full Paper

Bibliographic reference.  Irtza, Saad / Sethu, Vidhyasaharan / Le, Phu Ngoc / Ambikairajah, Eliathamby / Li, Haizhou (2015): "Phonemes frequency based PLLR dimensionality reduction for language recognition", In INTERSPEECH-2015, 997-1001.