To solve the problem in conversations made between plural master units and plural slave units of an intercom that when echo canceller parameters are successively learned by training, it occurs that depending on an equipment configuration or the other party of conversion, learning does not converge even when training is executed, so that the effect of an echo canceller cannot always be exhibited.

An echo canceller is installed in each terminal. A slave unit contains a storage part in which fixed parameters such as echo suppressor operation conditions and microphone gains, etc. dependent on the presence of an external speaker and variable parameters such as filter coefficients of the echo canceller, other than the fixed parameters, which are dependent mainly on an ambient environment such as an indoor spatial cubic volume, etc. By executing learning successively using the variable parameters as initial values to alter the parameters, it is possible to realize the best acoustic echo and line echo cancellation, noise rejection and conversion switching in various combinations of conversion.