In the Stereophonic Acoustic Echo Cancellation (SAEC) analysis, an adaptive filter is a necessary part that has a large number of weights or taps. The SAEC is depended on the room impulse response (RIR) and acoustic path where the cancellation is performed. However, a large tap-length results in slow convergence and increases the complexity of the tapped delay line structure for FIR adaptive filters. To overcome this problem, there is a need for an optimum tap-length-estimation algorithm that provides better convergence for the adaptive filters used in SAEC. This paper presents a solution to the problem of balancing convergence and steady-state performance of long length adaptive filters used for SAEC by proposing a Modified Recurrent Neural Network (MRNN) based tap-length-optimization algorithm. For the tap-length optimization analysis MRNN is developed, which is one of the Artificial Intelligence (AI) technique, it provides accurate results and to increase the convergence rate. The tap-length optimization is applied to a single long adaptive filter with thousands of coefficients to decrease the total number of weights, which in turn reduces the computational load. The proposed tap-length-optimization algorithm is applied to an existing multiple sub-filter-based echo canceller, for which present a convergence analysis. The proposed method is implemented in the MATLAB platform and the optimal output performances are demonstrated. The performances of the Mean Square Error (MSE) and Echo Return Loss Enhancement (ERLE) are determined. The proposed Method is compared with the existing methods like Artificial Neural Network (ANN) and Fuzzy ogic Controller (FLC).