Abstract: What role the window function plays in time-frequency analysis and the shortcoming of S transform are discussed. It is revealed that the result of the analysis is affected greatly if the window is narrow or the window duration is adjusted according to the analysis frequency. Based on STFT, a new method of time-frequency analysis is presented, where the window duration is adjusted automatically according to the time-varying characteristic of signal. Taking the change of instantaneous frequency as the criterion for the change of signal spectrum structure, the method can increase window duration before the change of instantaneous frequency so as to cause the window to cover the whole or majority of local stationary signal, thus obtaining the optimum time-frequency resolution for all local of stationary signal. The method proposed applies to analyzing and comparing the chirp signal, time-varying signal and the noisy overlapped signal. The results show that it can exactly represent the time-frequency distribution of signal with optimum time-frequency resolution obtained, and is non-sensitive to noise. The method is therefore regarded as advanced and practical.