A noise cancellation system with an improved performance and low computational costs is presented
in this paper. In speech applications, slow convergence and high computational burden are the main problems
incorporating with conventional noise cancellation method. The proposed noise canceller is based on using
multirate filter bank to split the spectrum of the input signals and uses the least mean square (LMS) algorithm
in branches to control a finite impulse response (FIR) filter to reduce the noise in the input noisy speech. The
computational power is greatly reduced by polyphase implementation and the noble identities. Direct and
polyphase models were developed, tested and compared to the equivalent full band model. The proposed
scheme shows better convergence behavior compared to classical approach with 50% reduction in
computational complexity.