Labeling a speech database with landmarks and features

Abstract

A database of utterances has been labeled for acoustically salient landmarks and for phonetic features. The landmarks locate syllabic nuclei, glide midpoints, and acoustic discontinuities at the narrowing or releasing of consonantal constrictions. The consonantal landmarks are labeled with values of the articulator-free features (e.g., [continuant] and [sonorant]), and values of articulator-bound features are recorded at each landmark based on acoustic evidence near the landmark. These features specify, for example, the major articulator and its placement, as well as the activity of secondary articulators such as the glottis and the soft palate. Conventions have been developed for labeling reduced vowels, syllabic consonants, glottal stops, consonant sequences in which closures or releases are not acoustically marked, etc. The labeled database is being used for two purposes: (1) to provide a documentation of instances in which the acoustic/articulatory representation of an utterance deviates from the underlying feature-based lexical representation and (2) to train and to test signal-processing algorithms that are being developed to locate the landmarks and to identify the features attached to these landmarks. An advantage of this labeling method is that it gives a more fine-grained phonologically-relevant representation of an utterance than a conventional labeling in terms of phonetic symbols. [Supported in part by NIH Grant No. DC02978.]