LEAST-SQUARES ERROR BEAMFORMING USING MINIMUM STATISTICS AND MULTICHANNEL FREQUENCY-DOMAIN ADAPTIVE FILTERING
Robert Aichner, Wolfgang Herbordt, Herbert Buchner, Walter Kellermann · 2003
In this paper we introduce a novel adaptive beamformer which also copes with incoherent background noise. After derivation of the optimum filter based on a weighted time-domain least-squares error criterion we present an efficient realization by applying a multichannel frequency-domain algorithm exhibiting RLS-like convergence. For the computation of this algorithm a simultaneous estimation of the power spectral density matrices of both, the noise signal and the noisy speech signal is necessary. Hence, we propose to use a novel approach based on minimum statistics to achieve this simultaneous estimation. Furthermore, the necessary estimate of a desired signal is generated by using single-channel spectral subtraction. The musical noise is avoided in our approach due to the inherent temporal and spatial averaging of our proposed beamformer. Experimental results show that the algorithm is well-suited for diffuse noise environments (e.g. car noise). Moreover, subjective listening tests confirm that a high speech quality can be obtained. 1.