Structured variability in acoustic realization: a corpus study of voice onset time in American English stops

Talkers differ greatly in the acoustic realization of speech sounds, a source of signal variation that must be overcome by human and machine listeners. The present study examined talker variability in voice onset time (VOT) across the six word-initial stop consonant categories (/ptkbdg/) of American English. Employing a large corpus of productions from more than 100 speakers, we replicated previous findings of significant variation in overall and stop-specific VOT means. However, we also identified several statistical generalizations within and across phonetic patterns of individual talkers. Speaking rate accounted for a large portion of VOT variance, with talkers differing considerably in the strength of this relationship. Stop category means showed high pairwise correlations, particularly among /ptk/. Additionally, stop-specific means and variances were highly correlated. The structured variation present in VOT could be exploited by both listeners and automatic recognition systems to facilitate robust perceptual adaptation from limited exposure to novel talkers.