Short signal classification using set-membership identification with application to speech labeling

A set-estimation-based classification technique for parametric models was suggested by the authors. In the present work, classification performance of the method is improved by use of a more structured measure of distance between points representing classes and the set estimates of models to be classified. Moreover, an enhanced optimization procedure significantly improves the set-estimates upon which the decisions are based. The set-based algorithm is found to be particularly effective for classifying short signal frames for which the estimator variance is often large.