A system for enhancement of noisy speech comprises an input unit is configured to subdivide the spectrum of the input signal into a plurality of frequency sub-bands and to provide time-frequency coefficients X(k,m) for a sequence [X(k, m′−D+1) . . . X(k,m′)] of observable noisy signal samples for each of said frequency sub-bands, where k and m are frequency and time indices, respectively, and D is larger than 1. The system further comprises enhancement processing unit configured to receive X(k,m) and to provide enhanced time-frequency coefficients Ŝ(k, m), a storage for statistical model(s) of speech and for statistical model(s) of noise, and an optimizing unit configured to provide said enhanced time-frequency coefficients Ŝ(k,m) using said statistical model of speech and said statistical model of noise, while considering said sequence [X(k, m′−D+1) . . . X(k, m′)] of observable noisy signal samples. Thereby the enhancement processing unit is able to determine the enhanced time-frequency coefficients based on the time-frequency coefficients for each of said frequency sub-bands.