Multi-channel deconvolution using Pade approximation

A new approach of estimating acoustic transfer functions from multi-channel reverberant signals is proposed. The ratio of two transfer functions is approximated by a polynomial by estimating one reverberant signal from another using least mean square estimation. Each transfer function is then estimated using Pade approximation. Two criteria are used to search for the most appropriate pair of transfer functions among all pairs generated by Pade approximation. The dereverberated signal is derived by multi-channel inverse filtering using the multiple input/output inverse theorem (MINT). Simulation using Gaussian noise convolved with minimum phase transfer functions gave a reconstruction SNR of 66 dB. The approach is also applied to sub-band inverse filtering of reverberant speech recorded in a real room.