Linear prediction of speech with a least absolute error criterion

Most methods for linear prediction of speech are based on a least square error criterion (L2criterion). In this paper, it is shown that the robustness of a least absolute error criterion (L1criterion) is particularly well adapted to the analysis of voiced sounds. The well-known linear programming L1algorithms have two important drawbacks: the stability of the synthesis filter is not guaranteed, and the computation load is too heavy for real time analysis. All these problems are alleviated by the new L1algorithm proposed in this paper. Experiments show comparable efficiency with classical L2methods.