The major difficulties for speaker-independent recognition of letters in French, result from the acoustic likness of some words ("M" and "N", "B" and "D", "F" and "S", "J" and "G", etc.) and from the lack of informations to solve ambiguities. This particular vocabulary requires the encoding of the main informations taken into account for the acoustic and phonetic decoding of continuous speech. The techniques the more frequently used to build up such systems are based on pattern matching. There are various pattern-coding and pattern-matching methods (Vector Quantization, Hidden Markow Modelling, Neural Network, Dynamic Time Warping, etc.) [Aldelfeld 80], [Burton 85], [Jelinek 85], [Huang 88], [Bulot 89], but most of them do not use explicitly coded knowledge. We are proposing a system based on a set of acoustic, phonetic, phonologic and lexical knowledge, represented by rules in Prolog II [Meloni 86]. The informations represented are various and they allow the localization and the identification of acoustic and phonetic phenomena (patterns, grouping of patterns, events, properties, cues, features, phonemes, syllables, etc.). A score is associated to each phenomenon described when the latter is identified. Rules and control use these scores in many ways to assign valuations to complex phenomena and to sort the most probable hypotheses of recognition.