ROBUST NONRECURSIVE AR SPEECH ANALYSIS

In this paper a robust non-recursive algorithm for estimating the line ar prediction (LP) parameters of autoregressive (AR) speech signal mod el is proposed. Starting from Huber's robust M-estimation procedure, m inimizing the sum of appropriately weighted residuals, a two-step robu st LP procedure (RBLP) is derived. In the first step the Huber's conve x cost function is selected to give more weights to the bulk of smalle r residuals, while down-weighting the small portion of large residuals , and the Newton-type algorithm is used to minimize the adopted criter ion. The proposed algorithm takes into account the non-Gaussian nature of the excitation for voiced speech, being characterized by heavier t ails of the underlying distribution, which generates high-intensity si gnal realizations named outliers. The obtained estimates are used as a new start in the weighted least-squares procedure, based on a redesce nding function of the prediction residuals, which has to cut off the o utliers. The experiments on both synthesized and natural speech have s hown that the proposed two-step RBLP gives more efficient (less varian ce) and less biased estimates than the conventional LP algorithms, and a one-step RBLP based on a convex cost function.