Blind separation problem is discussed, when sources are not independent, but have spatial variance dependencies. Hyvarinen and Hurri (2003) proposed an algorithm which requires no assumption on distributions of sources and no parametric model of dependencies between components. In order to obtain semiparametric algorithms which give a consistent estimator regardless of the source densities and the dependency structure, we study estimating functions for this model by the statistical approach of Amari and Cardoso (1997). Unlike the ICA model, the maximum likelihood estimation is not a semiparametric method in this case. Therefore, we consider a class of estimating functions which contain the quasi maximum likelihood estimation of the ICA model and the nonstationary ICA algorithm by Pham and Cardoso (2000). By modifying the score function, we got an estimating function close to it and proposed semiparametric algorithms based on it. Our algorithms were compared to other BSS methods with several artificial examples and speech signals.
展开▼
