Abstract:
Owing to the flexibility of time-frequency resolution of wavelet transform,we present an improved voice activity detector(VAD) based on wavelet transform.Robust parameters in different scale and time resolution are computed for VAD decision,such as silence measure,stability measure of amplitude spectrum between adjacent frames,background noise measure of different frequency band,time-domain stability measure of scale 1.The silence measure is used to detect the existence of silence in the input frame.The stability measure of amplitude spectrum between adjacent frames is adopted to give a rough decision of the detection of stable noise based on the assumption that background noise is stable.If current input frame is noise,the energy of every frequency band is below the average of background noise energy threshold over long time.We divide the signal bandwidth into several scales by wavelet transform,and calculate the background noise measure of different scale.In low scale the input signal changes rapidly,and the variety of short time energy will be removed with long window.We calculate the mean square error of short time energy,and get time-domain stability measure from detail coefficients of scale 1.With these measures,we make the VAD decision.Compared with G.729 Annex B,the authors can detect the voice activity more accurately and reduce the ratio of speech clipping using the new algorithm.And the improved algorithm can achieve robust performance for different background noise,even in serious low signal-to-noise environment about 10dB.