Some Statistics on Vowel Formant Variability.

摘要

Acoustic phonetic studies of vowels often present average results with little of the information on formant variability which is essential for defining the sizes of acoustic phonetic categories and for evaluating the comparative significance of contextual and interspeaker effects. In this study, phonetics instruction tapes provided 828 steady-state tokens of 17 unrounded and 13 rounded non-nasalized, non-retroflexed vowels produced by a single female speaker. The formant frequencies were measured via the linear prediction method, and the most steady-state 108.8 ms of each token was located by an automatic algorithm. Formant analysis errors and wild samples were eliminated by visual inspection of scatter diagrams. The remaining 779 tokens (94.7% of the original tokens) result in standard deviations of between 13 and 87 Hz for F sub 1; 32 and 145 Hz for F sub 2; and 28 and 159 Hz for F sub 3. Except for some indications of multi-modality for cases of very high standard deviation, no discernable systematic relation seems to exist between formant variability and (1) articulatory features of the vowels, (2) nativeness of the vowels, (3) formant frequency, or (4) formant bandwidth. Although some of the standard deviations are larger than expected on the basis of earlier studies, certain vowel configurations are found to be highly reproducable. This finding is related to acoustic phonetic vowel classification.