Adaptive nonlinear predictive analysis for speech using a cascaded LMS-VSLMS predictor

Summary form only given. When we perform linear predictive analysis on speech signals, prediction errors are inducted. To suppress these errors, we often rely on nonlinear predictors. A nonlinear predictor, however, possesses disadvantages, such as high complexity, slow convergence, necessity of using a large number of data samples, etc. We propose an adaptive nonlinear predictor which has the structure of a cascade of an LMS predictor and a VSLMS (variable step LMS) predictor. Experiments were conducted on continuous speech and the proposed predictor provided superior prediction gains compared with the LMS and VSLMS predictors. Furthermore, we investigated the proposed predictor with iterative processing. As a result, we confirmed that the simple cascaded predictor provides sufficient convergence.