Abstract: A delay estimation method based on encoder-temporal modeling structure is proposed to estimate the delay of microphone signal relative to the far-end signal in acoustic echo cancellation. In the proposed method, the far-end signal and the microphone signal in the short-time Fourier transform domain are used as input features. High-dimensional features with phase information are extracted by an encoder composed of complex convolutional neural networks. The memory ability of recurrent neural network is used to learn the time delay relationship between two input signals. A mapping from signal to delay is constructed by the proposed method. The simulation results show that the proposed method has the following advantages over WebRTC-DE and GCC-PHAT: (1) the number of parameters and computational complexity of the model are not affected by the delay; (2) the convergence time and tracking time of delay estimation are effectively reduced; (3) smaller and more stable estimation error and standard deviation are achieved in the case of long reverberation time and double-talk. Experiments on adaptive echo cancellation cascaded with the proposed delay estimation module verify the effectiveness of the new method.