A Comparative Study of Noise Cancellation Using Least Mean Squares Adaptive Filter and Recurrent Neural Network Filter

Abstract

Algorithms in artificial neural networks (ANN) are evolving as better alternatives to conventional algorithms applied in various electrical engineering applications in general and signal processing applications in particular. Specifically, we focus on a special type of ANN called recurrent neural networks (RNN), which delivers superior performance on sequential data due to the presence of internal memory. In the present paper, we comparatively analyze the performance of RNN and least mean squares (LMS) adaptive filter on audio data for active noise cancellation. We use normalized mean squared error (NMSE) as performance measure for comparison. Furthermore, we also investigate the number of epochs for training and the time taken to give the desired output via numerical simulations. Our simulations show that RNN filter delivers better NMSE performance than conventional LMS filter.