Some effects of duration on vowel recognitionJames M. Hillenbranda)and Michael J. ClarkDepartment of Speech Pathology and Audiology, Western Michigan University, Kalamazoo, Michigan 49008Robert A. HoudeRIT Research Corporation, 125 Tech Park Drive, Rochester, New York 14623共Received 14 April 2000; accepted for publication 14 September 2000兲This study was designed to examine the role of duration in vowel perception by testing listeners onthe identification of CVC syllables generated at different durations. Test signals consisted ofsynthesized versions of 300 utterances selected from a large, multitalker database of /*V$/ syllables关Hillenbrand et al., J. Acoust. Soc. Am. 97, 3099–3111 共1995兲兴. Four versions of each utterancewere synthesized: 共1兲 an original duration set 共vowel duration matched to the original utterance兲, 共2兲a neutral duration set 共duration fixed at 272 ms, the grand mean across all vowels兲, 共3兲 a shortduration set 共duration fixed at 144 ms, two standard deviations below the mean兲, and 共4兲 a longduration set 共duration fixed at 400 ms, two standard deviations above the mean兲. Experiment 1 useda formant synthesizer, while a second experiment was an exact replication using a sinusoidalsynthesis method that represented the original vowel spectrum more precisely than the formantsynthesizer. Findings included 共1兲 duration had a small overall effect on vowel identity since thegreat majority of signals were identified correctly at their original durations and at all three altereddurations; 共2兲 despite the relatively small average effect of duration, some vowels, especially/Ä/-/Å/-/#/ and /,/-/}/, were significantly affected by duration; 共3兲 some vowel contrasts that differsystematically in duration, such as /{/-/(/, /É/-/)/, and /(/-/|/-/}/, were minimally affected by duration;共4兲 a simple pattern recognition model appears to be capable of accounting for several features ofthe listening test results, especially the greater influence of duration on some vowels than others; and共5兲 because a formant synthesizer does an imperfect job of representing the fine details of theoriginal vowel spectrum, results using the formant-synthesized signals led to a slight overestimateof the role of duration in vowel recognition, especially for the shortened vowels. © 2000Acoustical Society of America. 关S0001-4966共00兲03912-6兴PACS numbers: 43.71.An, 43.71.Es 关KRK兴I. INTRODUCTIONDuration has long been a key feature in the descriptionand analysis of vowels. The chief phonological question con-cerns whether duration should be considered a contrastive orredundant feature 共Chomsky and Halle, 1968; Roca andJohnson, 1999兲, and the main phonetic issues have been themeasurement of vowel durations under a variety of condi-tions and the study of duration as a cue in vowel perception.In this study we are neutral regarding the phonological ques-tion of whether length should be considered an intrinsic pho-nological vowel feature in English. Rather we assume theexistence of distinct vowel categories that contrast with oneanother in most phonetic contexts and that are produced withdifferent typical durations in American English. Our focus ison the role played by variations in vowel duration in therecognition of vowel identity. Specifically, we studied theperception of 300 /*V$/ syllables that were synthesized infour different ways: 共1兲 an original duration condition inwhich the duration of each vowel was matched as closely aspossible to that of the original utterance, 共2兲 a neutral dura-tion condition in which the synthesis control parameterswere linearly stretched or contracted to produce a fixedvowel duration of 272 ms 共the mean of all 300 utterances兲,共3兲 a short duration condition in which vowel duration wasfixed at 144 ms 共two standard deviations below the mean兲,and 共4兲 a long duration condition in which vowel durationwas fixed at 400 ms 共two standard deviations above themean兲.A. Measurement studiesThe central phonetic fact underlying this study is thewell-known observation that American English vowels differfrom one another in average duration. Of particular interestare the many pairs of spectrally similar vowels that differ induration, pairs such as /{/-/(/, /É/-/)/, /,/-/}/, /|/-/}/, /Ä/-/#/,and /Å/-/Ä/. Average vowel duration measurements for the 12vowel types used in the present experiment are summarizedin Fig. 1. The data from Crystal and House 共1988兲 and vanSanten 共1992兲 are from connected speech, while the Hillen-brand et al. 共1995兲 and Black 共1949兲 measurements are fromCVC syllables. There is, of course, quite a bit of variabilityin the absolute durations associated with each vowel typeacross the four studies, reflecting the differences in speechmaterial. As expected, the two connected speech studiesshow shorter average durations than the two studies usingcitation-form syllables. The longer durations in Hillenbrandet al. than in Black are related primarily to the use of a finalvoiced stop 共/*V$/兲 in Hillenbrand et al. as compared to afinal voiceless stop 共/#V!/兲 in Black 共House and Fairbanks,a兲Electronic mail: [email protected] 3013J. Acoust. Soc. Am. 108 (6), December 2000 0001-4966/2000/108(6)/3013/10/$17.00 © 2000 Acoustical Society of America1953兲. Despite these differences in absolute duration, how-ever, the four studies show rather similar patterns of durationdifferences across the vowel categories. Correlations amongthe six possible pairings of the four functions shown in Fig.1 ranged from 0.70 to 0.95, with an average of 0.84. As willbe seen later, the results of our synthesis experiments showthat listeners are tacitly aware of these differences in typicalduration and make some use of this knowledge in makingjudgments about vowel identity.B. Pattern recognition studiesThe role of duration in vowel identification has beenstudied indirectly through the use of pattern recognition ex-periments. In work of this type, a statistically based patternclassifier is used to determine the separability of vowelsbased on various combinations of acoustic measurements.For example, Zahorian and Jagharghi 共1993兲 used a discrimi-nant classifier to identify signals in a database of 2922 CVCsyllables formed from nine initial consonants, 11 vowels,and eight final consonants. Zahorian and Jagharghi’s maininterest was the comparison of two