We are presently developing an Acoustico-Phonetic Decoding system which uses a Prolog II rule base combined with various neural networks. The rules organize the learning process by choosing relevant examples from a data base of sound. Their essential role in recognition is to describe the structure of the sounds ; they select the input data for the networks from the signal, and interpret their output according to the context. Different strategies were used for localizing and identifying vowels, fricatives and occlusives (depending upon the acoustic features of each macro-class); several network architectures were tested in parallel.