In vector predictive coding of speech signal sequences, the speech vector sequences that is obtained after a classification of adjacent sample of source speech sequences, can be regarded as autoregressive vector sequences with stable convariance. The vector predictive coding algorithm with highly parallel processing ability, can be achieved with the application of the principle of orthogonal projection in Hilbert space. In comparison with traditional lattice methods, the vector predictive coding algorithm has obvious advantages of calculating complexity and storage.