Spectral subtraction(SS) has been used for speech recognition in noisy environments. The maximum likelihood of noise mixed speech is analytically studied, which leads to the reformalization of SS. - Not only the subtraction factor used in conventional SS, but also a novel multiplication factor with elliptic integrals is used. This new SS method can be used alone or together with the conventional SS. The speech recognition experiments use a principal microphone and a reference microphone. The experiments suggest the effectiveness of the mixed method.