An iterative method for differentiating between known resonances and uncharacterized baseline contributions in MR spectra is described. The method alternates parametric modeling, using a priori knowledge of spectral parameters, with non-parametric characterization of remaining signal components, using wavelet shrinkage and denoising. Rapid convergence of the iterative method is demonstrated, and examples are shown for analysis of simulated data and an in vivo 1H spectrum from the brain. Results show good separation between metabolite signals and strong baseline contributions.