Even if received signals are highly cross-correlated, echoes can be effectively cancelled and no psychoacoustical problems arise. A received signal x.sub.i (k) (where i=1, 2, . . . , N) and an additive signal a.sub.i (k) are added together, and the added output is used to drive a speaker i and input into an echo cancellation filter 405.sub.i. The received signal x.sub.i (k) and the additive signal a.sub.i (k) are input into adaptive filters 401.sub.i and 402.sub.i, respectively. The difference between the sum of the outputs from all the filters 401.sub.i and all the filters 402.sub.i and an echo y.sub.m (k) is detected as an error e.sub.m (k). The coefficients of all the filters 401.sub.i and 402.sub.i are updated to reduce the error e.sub.m (k). When the error e.sub.m (k) is made sufficiently small, the coefficients of the filters 402.sub.i are transferred to the filters 405.sub.i. The sum of the outputs from all the filters 405.sub.i is detected as an echo replica, and the difference between the echo replica and the echo y.sub.m (k) is output.