Comparative Analysis of Adaptive Filters for Removal of Pink Noise from a Corrupted Speech Signal

Abstract

Noise affects different communication systems during transmission, on channels or reception processes and hence signal quality improvement is required when it is degraded due to various background noises. In this thesis work, Least Mean Square (LMS), Recursive Least Square (RLS), Wiener and Kalman filters are compared for removal of pink noise from a corrupted speech signal to improve some speech qualities using filter length, Signal to Noise Ratio (SNR) and Mean Square Error (MSE), computational complexity, stability and convergence speed parameters. A pure speech signal and pink noise are generated separately, added together and produce a noisy speech signal having different signal to noise ratio levels and then feed to the adaptive filters as an input. The filters then estimate the distorted speech signal and produce a mean square error that has a significant difference for the same input noisy signal. Based on the simulation results obtained, it is concluded that Kalman filter has better MSE performance in terms of filter length, signal to noise ratio and mean square error metrices, since it produces the smallest mean square error followed by wiener, LMS and RLS filters. In terms of computational complexity, stability and convergence speed metrices, Kalman is computationally more complex and has faster convergence rate but LMS is more stable. Hence, we can conclude that, in removing pink noise from a corrupted speech signal Kalman filter has better MSE and faster convergence speed performances even if it is computationally more complex and less stable.