For the control of lactic acid batch fermentations, the prediction of pH over a long time horizon and the determination of the final fermentation time are very useful information. Fermented milks with pure and mixed cultures of lactic acid bacteria are modelled with a feedforward neural network. For isothermal fermentations (42-degrees-C), a static model is sufficient to define the one and only reference curve. This reference fermentation combined with four sliding geometrical methods is used to perform a comparison with the actual fermentation. The final fermentation time, which occurs at a predetermined pH, is predicted with an accuracy of less than 20%. For fermentations conducted at various temperatures, a dynamical model is obtained as a recurrent neural network. Small order architectures are tested, and directed and semi-directed learning procedures are evaluated. The advantages of the semi-direted mode are shown and the resulting neural net is capable of predicting the final fermentation time with a mean relative error of 7.7% in the temperature range of 40-degrees-C-48-degrees-C.