Abstract: An effective method to smooth prior SNR is proposed. Based on the definition of prior SNR, this method smoothes prior SNR by using wavelet thresholding multi-taper estimators to reduce variance of speech and noise power spectrum.The method for smoothing prior SNR can avoid delay in the estimated prior SNR when energy level at the beginning of speech is high by using the decisiondirected method. It can also enhance estimation accuracy of prior SNR. Speech enhancement combined with this method can effectively suppress “music noise”, especially at low input SNR

Key words: prior SNR, spectrum estimation, speech enhancement