A number of various experimental studies has demonstrated that speech is a skilled serialorder motor task that is adapted to achieve time series of goals within a timing that does not allow any online cortical processing of feedback signals. In addition, the speech motor system is highly redundant with many available degrees of freedom, which makes the inference of motor commands from the physical signals an "ill-posed" inverse problem. These degrees of freedom are used in different ways to deal with sequence planning and variations in the conditions of articulation (various speaking rates, speaking while running, speaking while eating . .). For instance, clear anticipatory behaviors have been observed in articulatory and acoustic patterns associated with speech sequences. To deal with this complexity, speech motor control models classically consider a feedforward control scheme coupled with a feedback controller enabling a correction of motor commands in case of a wrong inference or in presence of an external perturbation. In this context, speech planning, which aims at solving the "ill-posed" inverse problem by setting the motor command patterns adapted to the production of a speech sequence, has been classically modeled within an optimal motor control framework. This approach has proven to generate results in close agreement with experimental data, in particular in terms of adaptation to perturbations or in terms of anticipatory behavior. However, criticisms of this approach include key issues, such as the nature of the neurophysiological mechanisms likely to be associated with cost computation and cost minimization, or the inability to account for the well-known token-to-token speech variability. This last drawback is inherent to the feedforward optimal control scheme, since it basically cancels possible variations along the degrees of freedom directions, by specifying a unique optimal solution to the control problem. In the present work, we propose an alternative approach by formulating feedforward optimal control in a Bayesian modeling framework. The methodology is based on the Bayesian Programming framework, which proposes a structure for the construction of Bayesian models. We consider this approach to be appropriate for solving the ill-posed problem while accounting for the observed token-to-token variability in a principled way, and preserving the basic principles underlying the search for optimality without being explicitly driven by the minimization of a cost. The approach is illustrated by reformulating GEPPETO, an existing optimal control model for speech production planning developed in the lab, into the Bayesian modeling framework. We demonstrate that models are nested, with optimal control as a special case of the Bayesian model: indeed, the Bayesian model reduces to the optimal control model, when the inferred control commands are strictly limited to those having the maximum posterior probability. Variability is formally generated by assuming that control is performed by sampling motor commands randomly according to the distribution solving the inference problem in the Bayesian model. An additional interest of the proposed formalism is its coherence for treating perception and action in a unified framework, as well as for dealing with additional constraints.