Phonetic labeling by template matching

Abstract

Results are reported of recent investigations into improving the basic template match scheme often used for initial phonetic labeling of speech segments. The technique is sensitive to (1) The set of templates (What are relevant phonetic classes?), (2) The manner of template derivation (Average a set of samples? Which set? Best representative? Most common?), and (3) The a priori labeling of training samples. (What is “correct” in Phonetics?) A novel approach to problems 1 and 3 is taken by employing the Harpy speech recognition system [B. Lowerre, Ph.D. dissertation, Carnegie-Mellon University, 1975] to define the best fit of a phonological dictionary to the segmental pattern space signal. Treating the population of training patterns as a distribution of (possibly mislabled) samples offers some solutions to (1) and (2) above. Intra-population clustering, using sample distributions to reject mislabeled training samples, and correcting the template match score by training population statistics are techniques which were investigated. Improved methods for template derivation, sample rejection, and match score correction yield significant improvements in labeling accuracy and overall speech understanding system performance.