This paper deals with Bayesian inference in measurement error models with unknown error covariances. Our formulation covers heteroscedastic and homoscedastic models for replicated data. Both equation-error and no-equation-error models are included in our proposal. Resorting to data augmentation, we present a simulation-based framework using the Gibbs sampler. Model selection is also briefly discussed. Results from a simulation study are reported. We work out an illustrative example with a real data set on measurements of mineral element contents in pottery samples. Copyright © 2012 John Wiley & Sons, Ltd.