Convolutive independent component analysis by leave-one-out optimal kernel approximation

This work addresses on blind separation of convolutive mixtures of independent sources. The temporally convolutive structure is assumed to be composed of multiple mixing matrices, each corresponding to a time delay, collectively transforming a segment of consecutive source signals to form multichannel observations. As τ = 1, this problem reduces to linear independent component analysis. For arbitrary τ , we propose a new algorithm to estimate the unknown convolutive structure as well as independent sources. The proposed convolutive ICA algorithm is based on optimal kernel estimation and leave-one-out approximation operated under the mean-field-annealing process. We test the new algorithm with artificially created data and twomicrophone recordings of speech and musics. It is shown that the error between the estimated and given convolutive structures is significantly reduced for artificially created data and the human speech is well separated from the background musics for the twomicrophone recordings. The proposed new algorithm is empirically shown effective for blind separation of convolutive mixtures of independent sources. Keywords– Convolutive mixture, independent component analysis, blind separation, optimal kernel estimation, leave-oneout approximation, K-state transfer function, mean field annealing.