This work aims at improving the automatic speech recognition of confusable words like letters. To deal with that problem, we have proposed a new discrimination method based on acoustic knowledge and artificial neural networks. In this paper, we present a validation of this method for speaker-independent tests, on a big database, with several additive noises at different Signal to Noise Ratio. The influence of the training conditions is taken into account and the results obtained show a good robustness of this method.