Statistical relationships in distinctive feature models and acoustic-phonetic properties of English consonants

The relationship between segmental contrasts, often modeled as being composed of distinctive features, and actual acoustic-phonetic properties is complex and many-to-many. Contrasts may be cued by a multiple acoustic-phonetic properties, and acoustic-phonetic properties often cue multiple contrasts. This paper presents a hierarchical multivariate statistical model of the relationship between a suite of 11 acoustic measurements and various target feature systems. Measurements are taken from 10 repetitions of 16 English consonants by 20 native speakers of English in both onset and coda position in nonsense monosyllables. Target feature systems range from models with little generalization across the various segments to ones that fully cross distinctive features to specify all of the segments. The statistical model enables analysis of within-speaker and between-speaker sources of variability in consonant production, and constitutes a principled statistical method for comparing these different distinctive feature models. Also, the role of the statistical model as a baseline for work on relations between different segments in perceptual categorization will be outlined.