Independent component analysis and MLLR transforms for speaker identification

In this paper, we explore the use of Independent Component Analysis (ICA) and Principal Component Analysis (PCA) techniques to reduce the dimensionality of high-level LVCSR features and at the same time to enable modelling them with state-of-the-art techniques like Probabilistic Linear Discriminant Analysis or Pairwise Support Vector Machines (PSVM). The high-level features are the coefficients from Constrained Maximum-Likelihood Linear Regression (CMLLR) and Maximum-Likelihood Linear Regression (MLLR) transforms estimated in an Automatic Speech Recognition (ASR) system. We also compare a classical approach of modeling every speaker by a single SVM classifier with the recent state-of-the-art modelling techniques in Speaker Identification. We report performance of the systems and score-level combination with a current state-of-the-art acoustic i-vector system on the NIST SRE2010 dataset.