A Gaussian/Laplacian hybrid statistical voice activity detector for line spectral frequency-based speech coders

In this paper we introduce a voice activity detection (VAD) algorithm that is based on a two-state hidden Markov model. The observation layer of the proposed model, that contains the state conditional probability density functions, is a Gaussian-Laplacian hybrid. The proposed algorithm provides a false detection rate that is significantly lower than that of G. 729 Annex B VAD. Given that it works in the domain of ITU-T G.729 parameters, it requires a minimal additional cost for feature extraction.