Controlling Speech Distortion in Adaptive Frequency-Domain Principal Eigenvector Beamforming
Ernst Warsitz, Reinhold Haeb‐Umbach · 2006
Broadband adaptive beamformers, which use a narrowband SNR-maximization optimization criterion for noise reduction, typi-cally cause distortions of the desired speech signal at the beam-former output. In this paper two methods are investigated to con-trol the speech distortion by comparing the eigenvector beam-former with a maximum likelihood beamformer: One is an ana-lytic solution for the ideal case of absence of reverberation and the other one is a statistically motivated approach. We use the re-cently introduced gradient-ascent algorithm for adaptive princi-pal eigenvector beamforming and then normalize the filter coef-ficients by the proposed distortion control methods. Experimen-tal results in terms of the achievable SNR gain and a perceptual speech quality measure are given for the normalized eigenvector beamformer and are compared to standard beamforming meth-ods. 1.