A new non-linear regression model for formant trajectories in English monosyllables incorporating dual targets for vowels.

Abstract

A new non-linear regression model is proposed to characterize the formant trajectories of the vocalic portion of English CVC syllables in the data described by Hillenbrand et al. [J. Acoust. Soc. Am. 109, 748–763 (2001)]. The modeling framework builds on work of Broad and Clermont [J. Acoust. Soc. Ame. 81, 155–165 (1987)], wherein formant trajectories were modeled via three additive components: (1) a single vowel target, (2) an exponential approach (in time from onset) toward the vowel target from an initial consonant onset value, and (3) an exponential approach (in time from offset) toward the vowel target from a final consonant offset value. The new model extends this to allow for a dual specification (nucleus + offglide) of the vowel targets [T. Nearey and P. Assmann, J. Acoust. Soc. Am. 80, 1297–1308 (1986)]. Initial results suggests that inclusion of a second vowel target provides substantial reduction (about 40%, 23% and 7% respectively for F1, F2, and F3) of error variance on trajectories averaged across 12 speakers. More detailed statistical analyses of variants of the new model are underway and will be reported for both the data described above and that reported on by Assmann et al. [this meeting].