UNDERDETERMINED SOURCE SEPARATION BY ICA AND HOMOMORPHIC SIGNAL PROCESSING
Stefan Winter, Walter Kellermann, Hiroshi Sawada, Shoji Makino · 2006
Nearly all approaches for underdetermined blind source separa-tion (BSS) assume independent and identically distributed (i.i.d.) sources. They completely ignore the temporal structure of col-ored sources such as speech signals. Instead, we propose a multi-variate model based on a multivariate Gaussian distribution that is then used to determine an unmixing matrix for underdeter-mined BSS. Based on parameterization by cepstral coefficients we present a novel ICA-based cost function for estimating the speech-related parameters of the unmixing matrix. Experimen-tal results support the proposed approach. 1.