A local learning system for acoustic phonetic decoding of French voiceless plosives is described. A rule-base is worked out from both a labelled acoustic phonetic data-base and a statistical one.

Locally, rules using the concept of fuzziness are arrived at through: yielding automatically (based on statistics) restriction boundaries for each /iPFi membership function that is associated to a (parameter P, function F) pair, working out a synthetic, arborescent AND/OR rule... Globally, an automaton is defined for each phoneme whose states refer to a part of the rule-base. Operating within a decoding procedure, this system allows to define rules and to test the various parameter/function pairing combinations.