A finite-dimensional approximation to general discrete-time nonlinear filtering problems is provided. It consists in a direct approximation to the recursive Bayes formula, based on a Hermite polynomials expansion of the transition density of the signal process. The approximation is in the sense of convergence, in a suitable weighted norm, to the conditional density of the signal process given the observations. The choice of the norm is in turn made so as to guarantee also the convergence of the conditional moments as well as to allow the evaluation of an upper bound for the approximation error.