Objective analysis versus subjective assessment of vowels pronounced by deaf and normal-hearing children

Objective whole-spectrum and formant analyses have been performed on all 15 Dutch vowels pronounced in /C1VC2/ words by 24 deaf and 24 normal-hearing children, in order to develop a model of pronunciation quality for evaluating (deaf) speech; the results as obtained for adult males by Bakkum et al. [J. Acoust. Soc. Am. 94, 1989-2004 (1993)] have been verified and extended. Spectral representations of the vowels were created by determining the output levels of a bank of 16 filters (90-7200 Hz), with 1/3-oct bandwidths and logarithmic spacing of their center frequencies. Spectral differences agree well with subjective differences in pronunciation quality obtained from magnitude estimation and identification experiments. Spectral differences not related to pronunciation quality judgments arise as a consequence of physiological interspeaker differences and variation in fundamental frequency, but these differences can be compensated for by speaker-normalization and F0-compensation procedures. Using principal components analysis (PCA), the vowel spectra can be described by a limited number of dimensions, without losing much information; a description in a two-dimensional PCA subspace still agrees well with the subjective judgments and it also agrees with a description by the first two formants. The whole-spectrum approach provides a determinate, readily interpretable model of pronunciation quality for evaluating vowels. As a practical advantage, its computational requirements are modest and, in conjunction with PCA, the vowel dynamics can be visualized, which makes the approach suitable for vowel training and diagnostics.