| |
Abstract:
The method presented here determines articulatory parameters
from acoustics in continuous speech on the basis of a search of
an acoustic-articulatory codebook designed from the data obtained
from experiments through a magnetic sensory system. The codebook
consists of pairs of segmented acoustic and articulatory
parameters which are, respectively, represented by 30 LPC
cepstral coefficients and the Cartesian coordinate of nine points
on the articulator. The time interval of a segment is 160 msec.
Furthermore, the method can estimate articulatory parameters more
precisely using the phonemic information in an utterance. It
determines articulatory parameters by selecting the articulatory
code vector in the codebook which minimizes the weighted distance
measure of the segmental spectral the distance, the squared
distance between succeeding articulatory parameters, and the
squared distance between the position designated by the code
vector and the position of the articulator predicted by the
minimum-acceleration model for the phonemic sequence. Experiments
were conducted to evaluate the efficiency of both constraints in
determining articulatory parameters by comparing the estimated
and the observed articulatory parameters. The results show that
the rms error between the estimated and observed articulatory
parameter was about 1.65 mm on average, and that the articulatory
features for vowels and consonants are recovered well.
|
