Man-machine dialogue systems using speech recognizers allow a more friendly asynchronous interaction, in such a way that the speaker is not forced to wait the end of previous prompts or messages uttered by the machine. In order to achieve that, man-machine systems accessed through the telephone network require an echo canceller, for eliminating the interference of the echo of the machine messages; this echo in unavoidable, given the mismatch of the hybrid circuit. Echo cancellers used in man-machine dialogue systems have different requirements from those used in man to man transmission systems; moreover their implementation adds a significant computational overhead to recognizers, inducing some possible degradation in recognition performance as well. Here we describe our specific implementation of an echo canceller for speech recognizers, not too heavy from the computational point of view and exhibiting good recognition accuracy as well.