Noise reduction in speech using adaptive filtering I: Signal processing algorithms

Abstract

Adaptive filtering is employed in configurations to filter narrow‐band speech corrupted by noise. These configurations utilize either an independent sample of the input noise or rely on correlation properties of the speech to accomplish cancellation. Necessary constraints on the algorithms to retain and/or improve intelligibility for normal and hearing‐impaired populations are presented. Previous work with the LMS algorithm is compared to the performance resulting from new methods which address the problem created by widespread eigenvalues in the speech plus noise correlation matrix. It is shown that failure to deal with this problem in the single reference application results in a muffling resonant distortion of the speech spectrum. Spectral averages of the original and processed versions of the speech segments are compared to determine improvement. Auditory tests are presented in the companion paper.