In this paper, we present the problem of speech recognition in adverse environment, especially in the case of specific telephone noise. We show that basic noise cancellation methods are not adequate to adapt a system to various telephonic networks. So, we propose a method aiming to identify the telephonic network on which the call has been done, by the mean of a neural network, to obtain new information about acoustic observations vectors. An HMM speech recognition system takes advantage of this new information and reduce his error rate by 25 % in our application.