Most of the algorithms for speech enhancement are designed to improve the speech listening comfort. However the frequency spectrum character is destroyed seriously after the speech enhancement. To achieve better speech listening comfort with less frequency spectral damages, we present an improved signal subspace algorithm for speech enhancement. Compared with the traditional signal space method, the improved algorithm can decrease the Mel-frequency Cepstral Coefficients (MFCC) distance, an evaluation measure which means less frequency spectral damages to the voice and keep the voices’ intelligence at the same time. Besides, the method can enlarge the distance of the easily confused voices, which means the improvement of the voice recognition ratio. Thus we get the purpose of the speech enhancement. The improved algorithm is used in a speech recognition program and has a good performance.