Abstract: A speech enhancement method under very low signal-to-noise ratios is proposed. There are two characteristics in the method based on and compared with the spectral subtraction. First, the subtraction parameter is proposed based on the masking properties of human auditory and is self-adaptive. Second, Bark wavelet transform is used, which is more suitable to human ears' auditory system, to give the before-and-after voice speech sound analysis. By performing an objective and subjective test, it shows that two improvements are made in the very low signal-to-noise ratio speech compared to the spectral subtraction algorithms. (1) reduce the residual noise and background noise more effectively. (2) the enhanced speech is more clearness and intelligibility.