Acoustic correlates of word stress in American English

Abstract

Acoustic parameters that differentiate between primary stress and non-primary full vowels were determined using two-syllable real and novel words and specially constructed novel words with identical syllable compositions. The location of the high focal pitch accent within a declarative carrier phrase was varied using an innovative object naming task that allowed for a natural and spontaneous manipulation of phrase-level accentuation. Results from male native speakers of American English show that when the high focal pitch accent was on the novel word, vowel differences in pitch, intensity prominence, and amplitude of the first harmonic, H1 * (corrected for the effect of the vocal tract filter), accurately distinguished full vowel syllables carrying primary stress vs. non-primary stress. Acoustic parameters that correlated to word stress under all conditions tested were syllable duration, HI*-A3*, as a measurement of spectral tilt, and noise at high frequencies, determined by band-pass filtering the F3 region of the spectrum. Furthermore, the results indicate that word stress cues are augmented when the high focal pitch accent is on the target word.

(cont.) This became apparent after a formula was devised to correct for the masking effect of phrase-level accentuation on the spectral tilt measurement, Hi *-A3*. Perceptual experiments also show that male native speakers of American English utilized differences in syllable duration and spectral tilt, as controlled by the KLSYN88 parameters DU and TL, to assign prominence status to the syllables of a novel word embedded in a carrier phrase. Results from this study suggest that some correlates to word stress are produced in the laryngeal region and are due to vocal fold configuration. The model of word stress that emerges from this study has aspects that differ from other widely accepted models of prosody at the word level. The model can also be applied to improve the prosody of synthesized speech, as well as to improve machine recognition of speech.