In this paper, the first algorithm for learning hybrid Bayesian Networks with Gaussian mixture and Dirac mixture conditional densities from data given their structure is presented. The mixture densities to be learned allow for nonlinear dependencies between the variables and exact closedform inference. For learning the network's parameters, an incremental gradient ascent algorithm is derived. Analytic expressions for the partial derivatives and their combination with messages are presented. This hybrid approach subsumes the existing approach for purely discrete-valued networks and is applicable to partially observable networks, too. Its practicability is demonstrated by a reference example.