In this paper we propose a mechanism for learning the parameters of a model that constitutes the basis of the natural language component of a speech understanding system. The model defines a representation of the meaning of a sentence as a sequence of elemental semantic units. The sentence production mechanism, in this paradigm, is equivalent to a noisy channel whose input is the sequence of meaning units and whose output is a sequence of acoustic observations. The decoding (i.e., the understanding) is then formalized as the problem of finding the meaning given the acoustic representation. The automatic estimation of the model parameters is possible if a statistically significant set of sentence examples is available, and if each sentence is provided with the correct meaning. Unfortunately a database of sentences annotated with their meaning is not available at the moment. Instead we have a database, within the DARPA ATIS project [4], in which each sentence is given a correct answer. In this paper we discuss the problem of automating the training procedure and we give some experimental results.