Adaptive multichannel linear prediction based dereverberation in time-varying room environments

This paper proposes a real-time acoustic channel equalization method with an adaptive multi-channel linear prediction technique. The proposed method takes a theoretically perfect channel equalization algorithm that also solves the problems existed in the actual implementation stage. The linear-predictive multi-input equalization (LIME) is an appropriate attempt for performing blind dereverberation with assuring the theoretical basis, however, it requires a huge computational cost for calculating large dimension of covariance matrix and its inversion. The proposed equalizer is formed as the multichannel linear prediction (MLP) oriented structure with a new formula that is optimized to time-varying acoustical room environments. Experimental results under both artificially generated and real room environments with a conversational speech confirm the superiority of the proposed method.