Spectral Subband Centroids for Robust Speaker Identification using Marginalization-based Missing Feature Theory

Abstract

Until now, marginalization-based Missing Feature Theory (MFT) for speech classification has been limited to the use of Log Spectral Subband Energies (LSSEs) as features. These features are highly correlated, thus suboptimal for classification with diagonal-covariance Gaussian Mixture Models (GMMs), a common classifier in marginalization-based MFT. In this paper, we propose that Spectral Subband Centroids (SSCs) are more apt for marginalization-based MFT, as they are both decorrelated and spectrally local. Our results show that SSCs as features produce a more robust marginalization-based MFT, diagonal-covariance GMM-based, Automatic Speaker Identification (ASI) system than LSSEs as features, for at all tested SNR values (with Additive White Gaussian Noise (AWGN)). It is also shown that a fully-connected Deep Neural Network (DNN) can accurately estimate the Ideal Binary Mask (IBM) used for MFT.