In this paper, we present a newmethod to segment silence and speech for noisy con-dition. Conventional segmentation methods usuallylack in robustness under high background noise be-cause they are mostly dependent on amplitude or en-ergy of speech signal. For speech signal, the corre-lation between the neighbor frequency bands includ-ing most speech energy is high and little effected bynoise, but the correlation between the neighbor fre-quency bands of noise is low. So we employ the cross-correlation of neighbor sub-bands of signal to locatespeech and noise. We first performed the wavelettransform to denoise and further calculated the cross-correlation between the wavelet coefficients in two se-lected sub-bands and then used the standard deviationof cross-correlation coefficients to segment speech andnoise duration. The simulation and the result analysisshow that this method is efficient for the low-energyphonemes even in low signal-to-noise ratio, and theamount of computation is less.
展开▼