Improved MO-LRT VAD based on bispectra Gaussian model

Introduction: Voice activity detection (VAD) remains a challenging problem in speech processing and affects a number of applications including noise reduction for digital hearing aid devices, speech recognition systems and speech coding for discontinuous speech transmission (DTX) in mobile and IP networks. During the last decade, researchers have paid attention to the study of discriminative features for classification and noise robust decision rules. A highly cited work is the VAD proposed by Sohn et al. [1], which is based on the evaluation of a single feature vector likelihood ratio test (LRT) and assumes a Gaussian model for the noisy signal DFT coefficients. The proposed algorithm considers a generalised LRT involving multiple and independent observations of the bispectra. The experimental analysis shows significant improvements over standardised and recently reported VAD methods.