A method is developed for fitting theoretically predicted astronomical spectra to an observed spectrum. Using a hierarchical Bayesian principle, the method takes both systematic and statistical measurement errors into account, which has not been done before in the astronomical literature. The goal is to estimate fundamental stellar parameters and their associated uncertainties. The non-availability of a convenient deterministic relation between stellar parameters and the observed spectrum, combined with the computational complexities this entails, necessitate the curtailment of the continuous Bayesian model to a reduced model based on a grid of synthetic spectra. A criterion for model selection based on the so-called predictive squared error loss function is proposed, together with a measure for the goodness-of-fit between observed and synthetic spectra. The proposed method is applied to the infrared 2.38--2.60 \mic ISO-SWS data (Infrared Space Observatory - Short Wavelength Spectrometer) of the star $\alpha$ Bootis, yielding estimates for the stellar parameters: effective temperature \Teff = 4230 $\pm$ 83 K, gravity $\log$ g = 1.50 $\pm$ 0.15 dex, and metallicity [Fe/H] = $-0.30 \pm 0.21$ dex.