Abstract: By analyzing the difference of time delay estimation of acoustic signals in different frequency bands in complex environment, a multi-band expectation value maximization time delay estimation method is proposed. In order to ensure that there is no overlap between the frequency bands, the signal frequency band is divided independently. We calculate the respective generalized cross-correlation of each sub-band, and then establish the maximum likelihood model for each generalized cross-correlation. The iterative solution of transforming multi-dimensional optimization into one-dimensional optimization using the expectation value maximization algorithm is derived to obtain the optimal sub-band generalized cross-correlation function, and the time delay information of the acoustic signal is estimated on this basis. The data simulation and actual experimental results show that this algorithm increases the percentage of the effective estimation value by 10% compared with the conventional delay estimation. In the application of steering response power, when the grid is 0.3 m. The obtained peak area convergence is more obvious, and the localization effect is better.