On the Performance of a Data-Reuse Fast RLS Algorithm for Acoustic Echo Cancellation

The recursive least-squares (RLS) algorithm is an appealing choice for system identification problems, mostly due to its fast convergence. However, its high computational complexity limits the implementation of this algorithm in applications like acoustic echo cancellation (AEC), where long length adaptive filters are needed. Nevertheless, the fast RLS (FRLS) algorithm represents a practical alternative for such challenging scenarios, since its computational amount is proportional to the filter length. At the same time, the FRLS algorithm becomes unstable especially for nonstationary inputs, like speech. To overcome this limitation, we present an improved version of the FRLS algorithm based on the data-reuse (DR) method. The resulting algorithm, namely DR-FRLS, is able to work using the maximum value of its forgetting factor (which leads to good stability), at the same time controlling its tracking behavior by tuning the data-reuse parameter. Simulations conducted in the framework of AEC support the good performance of the DR-FRLS algorithm, which could become a reliable candidate in real-world applications.