A geometric algorithm for voice activity detection in nonstationary Gaussian noise

A new algorithm for voice activity detection in additive nonstationary noise is presented. The algorithm utilizes the differences of the probability distribution properties of noise and speech signal. The Magnitude Density (MDF) and the Magnitude Distribution Functions (MDF) are defined. The noise level is monitored for automatic threshold estimation. The estimate is shown to be accurate also when analysis windows do not fully contain non-speech signals and in the presence of non-stationary noise. The algorithm has been applied different type of noises (traffic, water, restaurant, ect.). The voice activity detection algorithm is shown to operate reliably in SNRs down to 0 dB and noise variance up to 10 dB/sec.