A novel type of approximation scheme to the maximum likelihood (ML) approach is presented and discussed in the context of phylogenetic tree reconstruction from aligned DNA sequences. It is based on a parameterized approximation to the conditional distribution of hidden variables (related, e.g., to the sequences of unobserved branch point ancestors) given the observed data. A modified likelihood, based on the extended data, is then maximized with respect to the parameters of the model as well as to those involved in the approximation. With a suitable form of the approximations the proposed method allows for simpler updating of the parameters, at the cost of an increased parameter count and a slight decrease in performance. The method is tested on phylogenetic tree reconstruction from artificially generated sequences, and its performance is compared to that of ML, showing that the approach is competitive for reasonably similar sequences. The method is also applied to real DNA sequences from primates, yielding a result consistent with those obtained by other standard algorithms.