Dept. of Comput. Sci., Univ. of New South Wales, Australian Defence Force Acad., Canberra, ACT 2600, Australia
David J. Broad
Santa Barbara, CA 93105
A method is described for importing the concept of critical formant spacing into the cepstral domain, with a view toward automatic and robust classification of English vowels into back and front categories. The method is based on a cepstrum-to-formant, linear regression model reported previously [D. J. Broad and F. Clermont, J. Acoust. Soc. Am. 86, 2013--2017 (1989)], which is reformulated in order to transform low-order, linear-prediction (LP) cepstra into Bark distances between third (F[inf 3]) and second (F[inf 2]) formant frequencies. The back--front distinction itself is established using the critical 3.5-Bark distance as a binary decision criterion. The proposed method was evaluated through classification experiments involving (14th-order) LP cepstra and formant frequencies F[inf 2] (Barks) and F[inf 3] (Barks), which were all measured from three center frames of nine vowels spoken five times in /CVd/ context (C=/h,b,d,g,p,t,k/) by four adult, male speakers of Australian English. The accuracies obtained for correct back--front classification range between 94% and 100% both on an intra- and an interspeaker basis. These results confirm the effectiveness of relational measures of formant patterns for broad vowel distinction, but more importantly indicate the possibility of deriving certain phonetic features directly from the LP cepstrum.