We present a multi-window method for obtaining the time-frequency spectrum of non-stationary signals such as speech and music. This method is based on optimal combination of evolutionary spectra that are calculated by using multi-window Gabor expansion. The optimal weights are obtained by using a least square estimation method. An error criterion that is the squared distance between a reference time-frequency distribution and the combination of evolutionary spectra is minimized to determine the weights. Examples are given to illustrate the effectiveness of the proposed method.