First International Conference on Spoken Language Processing (ICSLP 90)
Coarticulation in continuous speech causes vowel formant tracks to be affected by nearby phonemes. Generally continuous speech causes the vowel formant targets to be centralised relative to their isolated word counterparts. The present study uses data from one male talker in the CSTR/ATR speech database saying 200 phonetically rich sentences. By concentrating on one talker, coarticulation can be studied without the confounding effects of accents, speech habits and formant ranges from many talkers. While some of these effects could be normalised out of multitalker data, some residual variance remains with all the methods we have tried. The 21 vowels of RP British English have been studied, the 14 monothongal vowels, /ii, i, a, e, aa, uh, oo, o, u, u", @@, @, 1/ and the 7 diphthongal vowels /ai, au, e@, ei, i@, oi, ou, u@/. Formant tracks were computed automatically and scanned for errors by hand. The formant frequency values at the right and left edges and the temporal center of the hand labelled vowel region are obtained. These values are plotted on scatter plots and analyzed using statistical methods. Generally the vowels are most effected by nearby semi-vowels /l, r, y, w/. No simple relationship between adjacent phoneme place of articulation and the vowel target change has been found. The data shows the presence of "robust vowels" which are not greatly effected by nearby semi-vowels. These vowels are not simply stressed vowels, but depend on duration and others factors being studied. Similar results were found in American English.
Bibliographic reference. Hieronymus, James L. (1990): "Preliminary study of vowel coarticulation in british English", In ICSLP-1990, 949-952.