Silence and speech segmentation for noisy speech using a wavelet based algorithm

In this paper, we present a new method to segment silence and speech for noisy condition. Conventional segmentation methods usually lack in robustness under high background noise because they are mostly dependent on amplitude or energy of speech signal. For speech signal, the correlation between the neighbor frequency bands including most speech energy is high and little effected by noise, but the correlation between the neighbor frequency bands of noise is low. So we employ the crosscorrelation of neighbor sub-bands of signal to locate speech and noise. We first performed the wavelet transform to denoise and further calculated the crosscorrelation between the wavelet coefficients in two selected sub-bands and then used the standard deviation of cross-correlation coefficients to segment speech and noise duration. The simulation and the result analysis show that this method is efficient for the low-energy phonemes even in low signal-to-noise ratio, and the amount of computation is less.