A Feature Normalisation Technique for PLLR based Language Identification Systems

Phone log-likelihood ratio (PLLR) features have been shown to be effective in language identification systems. However, PLLR feature distributions are bounded and this may contradict assumptions of Gaussianity and consequently lead to reduced language recognition rates. In this paper, we propose a feature normalisation technique for the PLLR feature space and demonstrate that it can outperform conventional normalisation and decorrelation techniques such as mean-variance normalisation, feature warping, discrete cosine transform and principal component analysis. Experimental results on the NIST LRE 2007 and the NIST LRE 2015 databases show that the proposed method outperforms other normalisation methods by at least 9.3% in terms of %Cavg. Finally, unlike PCA which needs to be estimated from all the training data, the proposed technique can be applied on each utterance independently.