关键词: 小波去噪, 阈值函数, 清浊音分离

Abstract: In low SNR，the traditional wavelet de-noising method may lose some useful ingredients of speech signal，resulting in low de-noising performance.As to the problem，this paper recommends a wavelet de-noising method by dynamic threshold based on separation of voiced signal and unvoiced signal.This method reduces a part of noise by spectral subtraction，and separates voiced signal from unvoiced signal with short-time energy.Wavelet package is used to protect the unvoiced signal.Avoiding to smooth speech signal，the method decides the dynamic threshold by the coefficients from each layer.A new threshold function is used to make up the disadvantages of soft-threshold and hard-threshold.By simulation testing，this method can improve the reconstructed quality of speech signal.

Key words: wavelet de-noising, threshold function, separation of voiced signal and unvoiced signal