Recently, I. Cohen has presented causal and noncausal algorithms to modify the classic decision-directed approach for prior SNR. It is well-known that prior SNR is critical to trade off the musical noise level and the audible clearness level in spectral subtraction speech enhancement. However, all these algorithms conflict with statistical signal model more or less. To adjust smoothing parameters which play an important role on the recursive procedure of prior SNR and noise spectrum estimate more reasonably, we present novel speech uncertainty state model which capitalizes on the masking property of auditory system, and propose a new modified approach which employs speech uncertainty probability to make automatic adaptation of smoothing parameters. Novel algorithm is capable of eliminating musical noise meanwhile lowering speech distortion by remaining original speech in the case of inaudible noise under masking threshold. Experiments confirm that novel algorithm is superior to classic methods, particularly at low SNR environment.