Underdetermined source separation for colored sources

This contribution focuses on the source separation stage as important part of underdetermined blind source separation (BSS). So far nearly all approaches for underdetermined BSS assume independently, identically distributed (i.i.d.) sources. They completely ignore the redundancy that is in the temporal structure of colored sources like speech signals. Instead, we propose multivariate models based on the multivariate Student's t or multivariate Gaussian distribution and investigate their potential for underdetermined BSS. We provide a simple yet effective filter based on the sources' autocorrelations for recovering the sources as basis for further advances in underdetermined BSS. The challenge is estimating the filter coefficients blindly. The experimental results support the idea that source separation for underdetermined BSS can be reduced to the separation of their autocorrelations.