To provide a voice encoding method which has improved compressibility by performing predictive encoding.

An adding circuit 1a calculates the sum signal (L+R) of signals L and R of two stereophonic channels and a subtracting circuit 1b calculates the difference signal (L-R). Difference arithmetic circuits 11D1 and 11D2 calculate the current and last differences Δ(L+R) and Δ(L-R). Predictive encoding circuits (15D1, 15D2, 16D1, 16D2) calculate a plurality of predicted values of the differences Δ(L+R) and Δ(L-R), calculate respective predicted residues of the plurality of the predicted values and of differences Δ(L+R) and Δ(L-R), and select the minimum predicted residue to perform predictive encoding.