Inverse truncated mixing matrix (ITMM) algorithm application to underdetermined convolutive blind speech sources separation

Inverse Truncated Mixing Matrix (ITMM) is a powerful method for underdetermined instantaneous blind source separation [1]. In this paper, we generalize ITMM algorithm to underdetermined convolutive blind source separation case. The proposed algorithm can be divided into two steps. The first step is the mixing filters estimation. The convolutive mixture can become an instantaneous mixture in time-frequency (TF) domain under some narrowband assumptions. Then, we used cluster method to estimate mixing matrix in every frequency bin. The second step is the source recovery part, we used ITMM method to mixing matrix in every frequency bin to source recovery in TF domain. Experimental evaluations are gained in artificial Room Impulse Responses (RIRs) environments, compared with conventional algorithms, the ITMM algorithm can separate speech sources to a higher signal-to-interference ratio (SIR).