In the Keal System [MER 89], as in most analytical speech recognition systems, the speech signal is processed in a bottom-up sequence of processing units. In Keal, these units are designed as follows: acoustic-phonetic decoding, lexical analysis, syntactic and semantic analysis. The purpose of this paper is to simplify the result of the first unit, namely the phonetic lattice, by removing its internal redundancy. This is achieved by using a connexionist network as a "filter" in order to tranform the lattice into a string. The task of the lexical decoder will henceforth be easier.