A speech enhancement algorithm based on non-linear filtering and noise masking

In this paper a concise form of noise masking is brought forward to serve as the speech enhancement criterion, which is used to estimate the parameter of a non-linear filter adaptively. With this parameter-adaptive non-linear filter, a computation-efficient speech enhancement algorithm is proposed. The efficiency comes from the coarse estimation of short-time power spectra and simplified calculation of masking threshold of speech. Finally the proposed algorithm is simulated and compared with the generalized spectral subtraction method and minimum mean-square error log-spectral amplitude estimation method for various noise types and levels with both Chinese and English speech segments. Objective measures and informal subjective listening tests show that the speech enhanced with the proposed algorithm has least audible residual noise.