This paper presents a unified approach to the problem of blind separation of sources, based on the concept of mutual information. This concept is applied to the whole source sequences as stationary processes and thus provides a universal contrast applicable to both the instantaneous and convolutive mixture cases. For practical implementation, we introduce several degraded forms of this contrast, computable from a finite-dimensional distribution of the reconstructed source processes only. From them, we derive several sets of estimating equations, generalizing those considered earlier.