A robust dereverberation technique for real-time hands-free speech recognition application is proposed. Real-time implementation is made possible by avoiding time-consuming blind estimation. Instead, we use the impulse response by effectively identifying the late reflection components of it. Using this information, together with the concept of Spectral Subtraction (SS), we were able to remove the effects of the late reflection of the reverberant signal. After dereverberation, only the effects of the early component is left and used as input to the recognizer. In this method, multi-band SS is used in order to compensate for the error arising from approximation. We also introduced a training strategy to optimize the values of the multi-band coefficients to minimize the error.