A common problem encountered in telephone communication networks is the presence of echo, which produced when the signal passes through telephone and telecommunication channels. Removal of this echo requires precise knowledge of the channel, which may be time varying. This calls for adaptive estimation of the channel, which is characterized by time varying impulse response. In recent years, there has been a marked interest in the application of adaptive filtering to acoustic echo cancellation (AEC) where the impulse response involved is long. In the fullband adaptive process the echo channel estimated with an adaptive FIR filter requires about 1000's of coefficients for perfect practical echo cancellation. Thus, this large number of coefficients is computationally intensive, as well as displays very slow convergence. For cost sensitive subscriber devices and high quality voice communications there are three main aims: (1) minimum computational complexity, (2) maximum convergence speed and, (3) maximum echo cancellation. It has been found that the best technique for acoustic echo cancellation must involve subband decomposition with adaptive algorithm in each branch. In this work, a hybrid technique has proposed that treats all these problems and gives suitable tradeoffs among the above aims. This new technique uses a nonuniform tree subband adaptive filter (NTSAF) system that has showed a noticeable increase in the convergence speed to about 5 times the old technique, better echo attenuation to -40 dB, and large reduction in the computational complexity. Furthermore, using NTSAF structure with polyphase filter banks (NTPSAF) will highly reduce the analysis/synthesis filters coefficients, which will result with 50% less delay in the reconstructed output speech signal. The algorithm has been described theoretically then modeled under MATLAB 7 simulation program using built-in filters and real input signals.