Data compression of nonlinear time series using a hybrid linear/nonlinear predictor

This paper presents a hybrid ADPCM that combines linear and nonlinear predictors, so that the advantages of both predictors can be utilized. This method estimates the linear part of the observed signal by the linear predictor, and then compensates the linear prediction error by the database-based nonlinear predictor. We develop a database update procedure so that the database size is not monotonously increased and nonstationary signals can be treated. The hybrid ADPCM achieves faster processing speed than a single nonlinear ADPCM and better compression performance than a single linear ADPCM and a single nonlinear ADPCM.