A parametrization strategy for stochastic variational inference with Markov Gaussian processes is presented for state estimation of a physical system whose underlying dynamical equations are partially or completely unknown.