Compensating for vowel coarticulation in continuous speech recognition

Coarticulation alters the vowel formant characteristics in continuous speech. Studies of isolated monosyilables in the Iiterature suggest that some phonemes cause more severe distortions than others. The largest changes are caused by /r /, /1/, /w /. Unstressed vowels are most affected. Previous studies by Holmes (6] and by us (13] indicate that these effects are even larger for continuous speech. Vowel recognition algorithms which do not take context into account in continuous speech normally achieve correct recognition of approximately 75 % for the three top choices from the recognizer. By developing methods which explicitly model the phonetic context, higher Ievels of performance are expected to be achieved. An ongoing study is being made of all 16 of the American English vowels using a subset the DARPA acoustic-phonetic data base, a phonetically labeled 6300 sentence data base with 630 talkers. A subset of 7 vowels /iy/, /1/, /eh/, /ae/, /o/ and /u/ have been studied in all major contexts. The formant temporal patterns are being examined for phoneme triples and quintuples with a vowel in the center. These formant patterns are discussed along with some effects of stress and speaking rate. Introduction Early studies of vowels by Potter and Steinberg (1] and Peterson and Barney (2] used isolated monosyllabic words to provide canonical formant frequencies for each monothong vowel in American English. Given a neutral h_d context, the vowels dustered weil in the space of the first and second formant, across hundreds of talkers. Vowels were believed to be characterized by the steady state frequencies of the first two or three formants. This view was challenged by some perceptual studies by Fairbanksand Grubb (3] for American English and Fujimura and Ochiai (4] for Japanese. They showed that the steady state portions of the vowels excised from the words were less identifiable than the whole words. In some way vowel context is very important in disambiguating vowels. Strange et al (5] showed that by listening to just the transitions into and out of the vowels with silence replacing the usual vowel, listeners were able to identify the vowel almost as weil as when played the whole word. An important cue to identification in this experimentwas duration. With duration not available as a cue, the vowel-less vowel identification accuracy feil to somewhere between isolated vowels and whole word. An early study by Shearme and Holmes (6] showed that vowels in continuous speech very seldom had steady states and often did not overlap the Peterson-Barney contours in any part of their frequency trajectories in time. Generally the vowels are much more centralized in continuous speech and the vowel formant regions overlap considerably due to coarticulation. The durations of vowels are Ionger for isolated words than for continuous speech. Vowels in front of pauses are also lengthened. One possible hypothesis from this data is that vowels in continuous speech are more identifiable than vowels in words bounded by silence. Presumably this is because a great deal of information is carried in the formant transitions. Strange et al (5] have advanced the theory that vowels are completely determined by the formant transitions into and out of the vowel. Nearey and Assmann (7] have postulated three alternate possible explanations for vowel perception, dual formant target, target plus slope and target plus direction. In their experiments they were unable to determine which mechanism was a clear best choice. Experiments of Gay [8] would seem to favor target plus slope, those of Pols [9] favor target plus direction for Dutch diphthongs. All of these studies were conducted on isolated words. The approach presented here is to attempt speaker independent vowel recognition based on formant frequencies and pitch. Speaker normalization using formant differences and F1 pitch as suggested by Syrdal and Gopal (10] has been studied for this data. The second source of variation * Institute for Computer Seiences and Technology, Building 225 Room A216, U. S. National Bureau of Standards, Gaithersburg, Md. 20899, U .S.A.