The Research Of Voice Activity Detection Method In Noise Environment

With the rapid development of information technology, the demands for finding a robust Voice Activity Detection, which has a good effect in human-computer interaction under noisy environment, becomes a hot topic. This thesis proposed a robust Voice activity detection based on noise classification and provided a new point of view for VAD in noisy environment.The Noise92 database is used to add noise to the speech data of Timit database to produce observed signals in different SNR and noise types. This paper analysis four kinds of VAD algorithms in fifty kinds of noise environment, and establish three kinds of classification methods based on the three eigenvalues.With the research of four kinds of traditional VAD method, we find that features of noise affect the accuracy rate. This thesis implements a robust activity detection algorithm that can switch the calculation method automatically depending on entropy of the noise which use spectral summation while entorpy of noise is higher and use spectral product at lower entropy condition. This thesis presents a developed long-term spectral divergence(LTSD) algorithm in sub-band using a noise estimation method which is updated by averaging the power spectrum of noisy speech using time and frequency dependent smoothing factors. This thesis proposes a developed PARADE(periodic component to aperiodic component ratio based Activity Detection) using a fused pitch. Three kinds of feather have been used to make classification, and uses the three-layered neural network to design a voice activity detection method. The method can switch algorithms depending on noise classification which is a compromise between the correct rate and computational complexity. The method selects the simplest algorithm of dual-threshold endpoint detection using short-time average zero-crossing rate and short-time average energy while the SNR is higher than 15 d B. If the SNR is lower than 15 d B, the method uses the classification method to judge the type of noise and choose the corresponding algorithm. The method selects the spectral summation algorithm under stationary and wide-band noise, spectral product based algorithm under stationary and narrow-band noise, developed LTSD algorithm under non-stationary noise contain pitch and developed PAR algorithm under non-stationary noise without pitch.The experimental results show that the proposed method can switch the algorithms using the classification method depending on the noise and achieves an average accuracy of 70%.