Adaptive nonlinear predictor based on order statistics for speech signals and its performance improvement by iterative approach.

In this paper, we propose a new adaptive filter and adaptive nonlinear predictor to implement the nonlinear prediction of speech signals to solve the problem of impulse-shaped residual errors in the linear prediction. While using order statistics processing, we add an improvement that keeps time information about the input signal to the LMS-L filter proposed by Pitas and Venetsanopoulos. Tests using synthesized speech show that the proposed nonlinear predictor has superior prediction performance to the Volterra series predictor and not just the linear predictor. However, even with this proposed nonlinear predictor, the prediction accuracy degrades in experiments using real speech. Therefore, we propose a new iterative method to conquer this problem. The iterative method pays attention to the periodicity of the speech, reuses the speech data, and obtains a prediction accuracy similar to when using many data samples. In experiments on real continuous speech, the proposed predictor improved prediction accuracy by using the iterative method, and the effectiveness of the method was confirmed. © 2005 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 88(11): 28–42, 2005; Published online in Wiley InterScience (<URL>www.interscience.wiley.com</URL>). DOI 10.1002/ecjc.20148 [ABSTRACT FROM AUTHOR]