Analysis and application of adaptive noise reduction using sparse filters

Summary form only given. The analysis of a sparse adaptive filtering technique and its application to the problem of system identification and noise reduction are discussed. In conventional adaptive filtering, modeling of systems whose impulse responses have clusters of nonzero coefficients, separated by samples that are small or zero, requires that the adaptive filter be sufficiently long to match the system. Consequently, in its converged state, the adaptive filter has many impulse response samples which are close to zero. These small coefficients contribute to residual filter misadjustment. Additionally, the convergence rate of the filter is determined by the total length. A sparse method that circumvents these problems by avoiding the calculations associated with the near-zero coefficients has been developed. As a result, the final mean square error attained is reduced, as is the convergence time.<>