Dereverberation Using  Multi Channel Blind System Identification  

Multichannel blind system identification (MCBSI) becomes a major methods for dereverberation. The main problem that leads to low accuracy of the identification in {MCBSI} is the common zeros. It is known that the probability of common zeros increases as the system order increases. Therefore there is a strong motivation to decrease the system order. A brief survey is provided, along with a review of two of the main methods for subspace (SS) MCBSI. We introduce several solutions to the well-known gain ambiguity (GA) problem that rises in the subband approach for {SS-MCBSI}. These solutions are more robust and outperform existing methods. In addition, we introduce a new method that improves the estimation in the subband approach by using several shifted versions of filterbank. This method increases the estimation accuracy in the overlapping area of the subband analysis filters. Moreover, a new framework that breaks the system into time partitions is introduced. This method removes the requirement for the use of long FFT devices, reduces computational burden and decreases the quantization error of the FFT. Finally, a new adaptive MCBSI method is introduced. This is an adaptive method based on the multi-delay filter (MDF) framework.
 
* This research was carried out towards the M.Sc. Degree in Electrical Engineering at Bar-Ilan University, under the supervision of Prof. Sharon Gannot.