An improved spectral subtraction algorithm for suppressing late reverberation is presented. Existing spectral subtraction methods, process similarly all frames of the reverberant speech signal. This work introduces two criteria for the identification of the speech frames that do not contain significant late reverberation power. The subtraction is adequately relaxed in these frames in order to preserve the signal from unwanted distortions caused by overestimation errors. These novel relaxation criteria together with previous improvements on the conventional spectral subtraction algorithms (i.e. the introduction of a perceptually motivated non linear filtering technique) result in an improved estimation of the anechoic speech signal. Objective assessments in terms of Noise to Mask Ratio and Weighted Spectral Slope Distance are used to demonstrate the performance of the algorithm. The results indicate that the proposed method achieves sufficient late reverberation suppression with fewer transient and spectral distortions.