The whispered speech database is established and the acoustic characteristics of whispered speech is analyzed.An improved spectral subtraction method based on band-partitioning power spectral entropy is introduced.This method detects the noise and voice part by analyzing band-partitioning power spectral entropy of the whispered speech

and improved spectral subtractions are processed separately on the two parts in order to get better denoising effect.The experiment shows that the method can separate the noise and voice part of the whispered speech effectively and has better enhancement effect on the whisper compared to the traditional spectral methods.