A Normalized Kurtosis based Voice Activity Detection Method under Non-Gaussian Non-stationary Noise

In multimedia systems, the performance of speech signal processing is often degraded by non-Gaussian noises such as applause, cough. This paper proposes a voice activity detection(VAD) method based on higher order statistics of linear prediction residual under non-Gaussian noises. This method utilizes the normalized Kurtosis statistics of linear prediction residual(NNK)as the determination criterion to perform VAD detection, because the NNK criterion can characterize the numbers of harmonics of the signal and there is apparent difference between the number of harmonics speech and non-Gaussian noise. The algorithm trace the non-stationary Gaussian noise by iteratively updating SNR. The simulation results show that the proposed method can effectively distinguish speech from non-Gaussian noise under non-stationary Gaussian noise.