Multichannel voice activity detection with spherically invariant sparse distributions

We propose a statistical multichannel voice activity detection algorithm by modeling the frequency components of speech signals as sparse multivariate complex distributions. In particular, we formulate a likelihood ratio test by modeling a multichannel speech observation as a spherically invariant random process with a parameter governing its sparseness. In addition, we consider reverberation as a component of the statistical model. Experimental results show that our proposed method significantly reduces false-alarm rate for reverberation tails and that sparse distributions provide higher detection accuracy compared to the traditional Gaussian distribution.