The invention provides a single-channel time-frequency overlapping signal blind separation method. The single-channel time-frequency overlapping signal blind separation method comprises the following steps: step 1, establishing a signal model and converting received mixed signals into a plurality of Gauss amplitude modulation source signals; step 2, calculating a joint maximum likelihood function of the mixed signals and changing the solving process of the mixed signals to a process of solving a multidimensional variable parameter; step 3, estimating a time domain center and a time width modulation parameter of each source signal base on signal characteristics; step 4, calculating an initial value of the multidimensional variable parameter; step 5, calculating an optimal value of the multidimensional variable parameter; and step 6, calculating the Gauss amplitude modulation source signals. The single-channel time-frequency overlapping signal blind separation method is a multidimensional variable parameter estimation method with a pre-estimated hybrid genetic-minimum value search algorithm; and the time width modulation parameter P of the source signal and the time domain center t0 of the source signal are firstly subjected to pre-estimation, and then the other parameters are calculated by using the genetic algorithm and the minimum value search algorithm. Therefore, the rate of convergence and the calculation accuracy are improved.