This paper presents an efficient approach to the single microphone speech separation problem based upon the idea of the Mix Max function. We show that the log amplitude of the short term Fourier transform of the mixed signal is almost exactly the maximum element wise of the log magnitude of STFT of individual signals by which the need for phase information whose characteristic are difficult to be parameterized or modeled is eliminated. Following that, we apply this function to construct a binary mask to separate each speech signal. The accuracy of the model is clarified through the use of the subjective and objective tests for different parts of overlapping speech segments
