An evaluation of a spectral moments metric with voiceless fricative obstruents

Abstract

The classification efficacy of a spectral moments metric was tested on a corpus of voiceless fricatives. The metric classified phonetic identity on the basis of mean, variance, skewness, and kurtosis values derived from cross-sectional spectra. The classification power of both linear- and Bark-based versions of the metric was tested using a corpus of 420 voiceless fricatives (/f/, /θ/, /s/, /∫/, /h/) obtained from multiple talkers and vowel environments. Discriminant function analyses performed on linear and Bark moment profiles of well-identified fricative tokens resulted in overall classification accuracies of 78% and 74%, respectively. Classification accuracies were substantially higher when the nonsibilant (/f/ and /θ/) data were excluded. In an attempt to improve metric classification of nonsibilant tokens, an experiment designed to identify the perceptually appropriate location of a fricative analysis window was performed. The results suggested that nonsibilant classification performance of the metric may be substantially improved when moment information is based upon the spectral information contained within the onset and/or offset (transition) portion of the frication. [Work supported by NINCDS Grant NS19653 and NIDCD Grant DC00219 to SUNY at Buffalo.]