An information theoretic approach to speech feature selection applied to speech detection
Hussein A. Magboub, Samuel T.V. Alexander · NCSU Libraries Repository (North Carolina State University Libraries) · 1983
The selection of speech waveform features for use in speech detection algorithms has often been approached from an heuristic or ad hoc viewpoint.This paper applies formal information theoretic concepts to the problem of optimal selection of these speech features.The mutual information conveyed about a classification (i.e., speech present or speech absent) by the measurement of specific features is used as the information metric.The classes of features examined include energy, first autocorrelation lag, zero crossings per frame, linear prediction error, and first adaptive linear prediciton (LP) coefficient.It is shown that, of these features, the first adaptive LP coefficient provides the most information about the speech/no speech classification.Additionally, the mutual information measure is used to categorize sets of two features according to their speech decision information content.Among the sets of two features examined, the first adaptive LP coefficient and energy are found to be the optimal set.The extension to higher order sets is straightforward.