On estimating length and shape of the vocal tract
Edward P. Neuburg · The Journal of the Acoustical Society of America · 1999
Automatic vowel recognition is sometimes done by trying to estimate vocal-tract areas, usually derived using linear predictive coding. In the derivation there is a free parameter, namely, the length of the vocal tract; every choice of length leads to a different set of areas. This paper describes a revisit to experiments done in the early 1970s, in which the chosen length is the one that makes the vocal-tract shape most ‘‘human.’’ Using some improved criteria for optimality of shape, involving both smoothness and small variance, it is found that: (a) estimates of vocal-tract length are consistent, or at least highly correlated, with actual length; and (b) the tube shape (area function) derived using the optimal length seems to give a reliable characterization of the vowel, independent of the talker (and, in particular, independent of the talker’s gender). Classification results are about the same as those using the currently popular cepstral coefficients, but areas have the possible advantage that they are physiological features that the talker can actually control.