To prevent a coefficient vector for a modulated wave candidate of repeated identical codes from diverging, an error calculating part (36) calculates the difference between a vector X(i) with elements of input signals x(i) at N successive times and a vector Y.sub.m (i) of N replica signals y.sub.m (i) received from a transversal filter (34) to obtain an error vector E.sub.m (i), the square of the norm of which is supplied to a maximum likelihood sequence estimating portion (31), when signal estimation is carried out. In addition, a code sequence candidate {a.sub.m } is produced and a corresponding modulated wave candidate s.sub.m is generated. The filter (34) calculates the inner product of the modulated wave candidate s.sub.m and the coefficient vector W.sub.m (i-1) corresponding to each state to produce y.sub.m (i). W.sub.m (i-1) of each state is updated to W.sub.m (i) using an inner product vector of a generalized inverse matrix produced from s.sub.m corresponding to the state transition selected by the estimating portion (31) and E.sub.m (i).