Two-Sided Filters for Frame-Based Prediction

A new linear prediction model, based on a two-sided predictor which predicts on the basis of past and future samples within a frame is presented. The new linear prediction model may be applied wherever frame-based prediction is employed. A stable synthesis procedure is derived by casting the prediction equation as a cyclic convolution in the time domain. When the filter order is the maximum possible, the synthesis filter is shown to have a frequency response proportional to the squared magnitude of the DFT of the frame. A symmetric two-sided predictor is described which has only half the number of coefficients to be coded as compared to a one-sided predictor of the same order. Two-sided prediction showed at least 5 dB improvement in prediction gain over one-sided prediction in our simulations on speech data. Whether this translates to coding gain will be known only after further studies with CELP-type coders. © 1991 IEEE