D-LTSV Voice Activity Detection Method Based on Dynamic Feature
Zhao Hua · Jisuanji gongcheng · 2014
Voice Activity Detection(VAD)is a critical step for speech processing. In order to improve the performance of VAD in low Signal-to-noise Ratio(SNR)and nonstationary noise,this paper proposes a novel D-Long-term Signal Variability(LTSV)method based on LTSV for VAD. It uses the Bartlett-Welch method to estimate the signal spectrum,analyzes the entropy on the signal spectrum,and utilizes the analytical method of dynamic features used in the cepstrum to extract dynamic features of the entropy. D-LTSV takes into account the degree of nonstationarity of the signal and the dynamic changes between the frames. Compared with LTSV,experimental result shows that D-LTSV owns more discriminative power which is improved by50.77 percent in low SNR and nonstationary noise and makes VAD more robust and more accurate.