A system for multi-speaker letter recognition in the French alphabet is presented. The main characterictic of this system is to combine a knowledge-based segmentation with a connexionist classifier. The segmentation module detects, automatically for each letter, discriminant acoustic patterns. For the classification module, two different approachs are tested. One is unsupervised using Self-organizing Feature Maps ; while the other requires supervised training data using Linear Vector Quantization classifier. A problem called bord effect appears with the first classifier and mainly explains the recognition rate. Best results are obtained with a new version of the second classifier which improves the recognition rate of 10%.