Different empirical formulas have been proposed to describe the water retention curve (WRC) and relative permeability (k_r) of soils. This paper presents a Bayesian framework that evaluates not only the most probable empirical fitting constants, but also their joint probability density function. A dataset containing two soil classes — sand and silty loam — compiled from the UNSODA database is used for illustration. First, model constants of the van Genuchten’s WRC formula are calibrated and subsequently used to predict k_r of the studied soils using two existing formulas based on Mualem’s and Burdine’s models. The best estimated k_r in both formulas is found to skew towards the lower side of the measurement. Then, a new three-parameter empirical formula is proposed to describe k_r with suction while the model constants are calibrated from the permeability data. Using the proposed framework, the statistical distribution of k_r and subsequently the unsaturated permeability (k_unsat), as a function of suction, can be obtained. The results are then applied to a hypothetical two-layer capillary barrier composed of soils of the compiled dataset to determine the breakthrough suction (ψ_BT) of the barrier. The proposed Bayesian approach gives a probabilistic distribution of ψ_BT instead of a single value in the traditional deterministic method.