Abstract:
Hillenbrand and Nearey (abstract 3aSC12) present results of a perceptual experiment involving the identification of 300 natural and 600 synthetic /hVd/ syllables of American English. Results of linear logistic regression of the same data are presented which confirm and clarify the basic findings of the companion paper. This analysis allows the optimized fitting of a generalized linear model to listeners' responses using stimulus properties as predictors [T. M. Nearey, J. Phon. 18, 347--373 (1990)]. Models including both nucleus and offglide formant measures (from steady-state and 80% of the total duration portions of the vocoids, respectively) reduce prediction error measures by about 50% over models including steady state measures only for dynamic formant stimuli. Evaluation of the contribution of individual formant and formant difference measures to specific vowel contrasts will be discussed. Comparisons will be made with results from linear discriminant analysis reported by Hillenbrand and Nearey.