Adaptive Filter Design and Implementation for Acoustic Noise Cancellation

Abstract

To minimize the noise level in speech signals, many adaptive

filters, such as Least Mean Square (LMS) and Normalized Least Mean Square (NLMS), are utilized to reach the steady state. The filter weights are adapted based on specific functions to enhance the Signal to Noise Ratio (SNR) of the system output with fast convergence speed. This thesis proposes a new LMS-based variable step size algo- rithm to reduce the noise level in the corrupted speech. The new algorithm is called Regularized Square Root Absolute Er- ror LMS (R-SRAE-LMS). R-SRAE-LMS switches between two different variable step size algorithms. The first algorithm is the approximated solution of the Regularized NLMS (RNLMS) which has the ability to reach the steady state very fast. The second algorithm is the Square Root LMS (SRLMS) which is able to achieve high output SNR value by eliminating the residual noise samples. This new algorithm is compared with other variable step size algorithms in speech enhancement. The results show that the new proposed algorithm improves the SNR at the filter output. It has fixed SNR improvement over a wide range of input SNR.

Over this range, the proposed algorithm has more stable perfor- mance measures and achieves the highest convergence speed and the lowest steady state error. It is also able to get very small misalignment values between the filter weights and the targeting channel. Moreover, this thesis proposes a detailed design and Field Pro- grammable Gate Array (FPGA) implementation of the proposed adaptive algorithm. The design of the adaptive filter is divided into a forward path and two feedback paths. The device uti- lization, operating frequency and the power consumption after a complete implementation process are also presented. It shows remarkable results compared to other variable step size based adaptive filter designs.