In this paper, we propose the introduction of a parametric weighting function to constrain the segmentation path which governs the training of a hybrid neuromarkovian scheme for an on-line word recognition system. Due to the parametric properties of this function, it is possible to modulate the constraints from imposing a strict balanced path with the same duration for every states to a totally free segmentation path. So far, during the initialization step of the training, when the neural network has little ability to correctly segment the word into its basic constituents, the constraints will be activated, and then, they will be relaxed to allow more flexibility to the segmentation-recognition process. Recognition experiments on the Ironoff database demonstrated that the proposed method allows to increase the word recognition rate when compared to a totally unconstrained segmentation training.