A Weighting Threshold Optimization Method in Speech Recognition

Speech recognition accuracy tends to drop down in strong noisy conditions. Accordingly, useful signals may be eliminated if the coefficients are processed with the same wavelet threshold. To solve this problem, a weighting threshold optimization method is proposed. The weighting is optimized according to the relativity of wavelet coefficients in each decomposed layer so that the denoising threshold can be more accurate. The useful speech signals can be reserved completely. The signals anti-interference ability can also be enhanced an get a better performance especially at adverse conditions.