PNCC-ivector-SRC based speaker verification

Eliathamby Ambikairajah, Jia Min Karen Kua, Vidhyasaharan Sethu, Haizhou Li

Abstract: Most conventional features used in speaker recognition are based on Mel Frequency Cepstral Coefficients (MFCC) or Perceptual Linear Prediction (PLP) coefficients. Recently, the Power Normalised Cepstral Coefficients (PNCC) which are computed based on auditory processing, have been proposed as an alternative feature to MFCC for robust speech recognition. The objective of this paper is to investigate the speaker verification performance of PNCC features with a Sparse Representation Classifier (SRC), using a mixture of ℓ <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf> and ℓ <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> norms. The paper also explores the score level fusion of both MFCC and PNCC i-vector based speaker verification systems. Evaluations on the NIST 2010 SRE extended database show that the fusion of MFCC-SRC and PNCC-SRC gave the best performance with a DCF of 0.4977. Further, cosine distance scoring (CDS) based systems were also investigated and the fusion of MFCC-CDS and PNCC-CDS presented an improvement in terms of EER, from a 3.99% EER baseline to 3.55%.