Voice activity detection in nonstationary Gaussian noise

The problem of voice activity detection in additive nonstationary Gaussian noise is considered. A new algorithm for the problem is presented. The algorithm utilizes the differences of the probability distribution properties of noise and speech signals. The magnitude density (mdf) and the magnitude distribution functions (MDF) are used to monitor the noise level for automatic threshold estimation. The estimate is shown to be accurate even when the analysis window does not fully contain non-speech signals and even in the presence of nonstationary noise. The voice activity detection algorithm is shown to operate reliably in SNR down to -5 dB. The method is compared with the periodicity measure method and zero-crossings method. Finally, a fusion algorithm utilizing those methods is suggested.