Voice activity detection based on a family of parametric distributions

Abstract

In this letter, generalized gamma distribution (GΓD) is introduced as a new statistical model of spectral distribution to be applied to the likelihood ratio test performed in voice activity detection (VAD). A gradient-based on-line algorithm is proposed to estimate the parameters of GΓD according to the maximum likelihood criterion. Experimental results show that the VAD algorithm implemented based on GΓD outperformed those adopting other parametric distributions. © 2007 Elsevier B.V. All rights reserved.