Phase-Processing For Voice Activity Detection: A Statistical Approach

Conventional voice activity detectors (VAD) mostly rely on the magnitude of the complex valued DFT spectral coefficients. In this paper, the circular variance of the Discrete Fourier transform (DFT) coefficients is investigated in terms of its ability to represent speech activity in noise. To this end we
model the circular variance as a random variable with different underlying distributions for the speech and the noise class. Based on this, we derive a binary hypothesis test relying only on the
circular variance estimated from the noisy speech. The experimental results show a reasonable VAD performance justifying that amplitude-independent information can characterize speech
in a convenient way.