The purpose of this work is to improve the automatic recognition of confusable words, considering such typical examples as French and American-English Alphabets. Our study proposes a comparison between global methods like DTW or HMM and a new method using neural networks. This method is based on the search for 2 discriminative frames inside the confusable words bearing the distinction between them. Then a parametrization is done and resulting vectors are given to neural networks. The tests conducted on normal speech, Lombard speech without additive noise and Lombard speech with additive noise show a general improvement of the recognition accuracy.