Gary Weismer
Michelle Molis
Paul Milenkovic
Waisman Ctr., Univ. of Wisconsin--Madison, 1500 Highland Ave., Madison, WI 53705
Karen Forrest
Indiana Univ., Bloomington, IN 47405
Several different schemes for extracting consistent spectral features of stop bursts have been proposed, with a fair amount of success in classifying place of articulation. Most of the material that has been used for classification tests has been of citation form, and in some cases carefully articulated citation form. The present experiment investigates the hypothesis that classification will be better in connected speech samples than in citation form. The hypothesis is motivated by the expectation of a diminished influence of vowel identify on stop bursts in the connected speech sample, where vowel reduction is likely to occur. The vowel reduction hypothesis has been confirmed by vowel duration and format frequency analyses; the classification analyses, based on discriminant function classification of spectral moments [Forrest et al., J. Acoust. Soc. Am. 84, 115--123 (1988)] are in progress and will be reported at the meeting.