On robust AR speech analysis based on quadratic classifier with heuristically decision threshold

The paper considers a robust recursive procedure for identifying a nonstationary AR speech model based on a quadratic classifier with a heuristic decision threshold. Two versions of the robust procedure with heuristic decision threshold, based on a frame-based quadratic classifier and a quadratic classifier with a sliding training data set, are evaluated and compared through analyzing natural speech signals with voiced and mixed excitation segments. The results obtained show that the considered robust procedure with the quadratic classifier with sliding training data set and heuristic decision threshold achieves more accurate AR speech parameter estimation, provides improved tracking performance, and achieves better discrimination capabilities for possible application in some vowel recognition systems.