Paper data
Title:
Nonlinear Prediction Based on Score Function Author(s): Jordi Solé i Casals, Signal Processing Group. University of Vic Enric MonteMoreno, TALP. Polytechnical University of Catalonia, Page numbers in the proceedings: Volume III pp 533536 Session: Non linear Speech Processing
Paper abstract
The linear prediction coding of speech is based in the assumption that the generation model is autoregresive. In this paper we propose a structure to cope with the nonlinear effects presents in the generation of the speech signal. This structure will consist of two stages, the first one will be a classical linear prediction filter, and the second one will model the residual signal by means of two nonlinearities between a linear filter. The coefficients of this filter are computed by means of a gradient search on the score function. This is done in order to deal with the fact that the probability distribution of the residual signal still is not gaussian. This fact is taken into account when the coefficients are computed by a ML estimate. The algorithm based on the minimization of a highorder statistics criterion, uses online estimation of the residue statistics and is based on blind deconvolution of Wiener systems [1]. Improvements in the experimental results with speech signals emphasize on the interest of this approach.
Paper
