Source separation, using a number of references that largely exceeds the number of sources, is hindered by the ill-condition of the cross-spectral matrix of the references. This has been solved in literature by reducing the dimension of the reference set, for instance by a truncated eigenvalue decomposition or Gram-Schmidt orthogonalization. The main drawback of these techniques is that the resultant virtual references are not necessarily maximally correlated with the sources. This paper exclusively introduces an original solution for this problem, coined "Maximally-Coherent Reference Technique", based on finding a set of virtual references so that they are maximally correlated with the sources. The number of sources is estimated using parametric bootstrap, which is advantageous over the asymptotic likelihood ratio test. The method is tested experimentally using numerical and real data, and is proved to solve the source separation problem while avoiding the shortcomings of the existing reference-based techniques.