Objective analysis versus subjective assessment of vowels pronounced by native, non-native, and deaf male speakers of Dutch

Abstract

An objective analysis has been performed on all 15 Dutch vowels pronounced in /hVt/ words by nine native Dutch, nine non-native, and six deaf males. Spectral representations of the vowel segments were created by determining the mean output levels of a bank of 16 filters (90-7200 Hz), with 1/3-oct bandwidths and logarithmic spacing of their center frequencies. The adequacy of the objective analysis is determined by the extent to which spectral information provides an accurate description of pronunciation quality. Spectral distances between the 24 utterances of each monophthong agree rather well with subjective distances obtained by listeners in an elaborate paired-comparisons experiment. For the various monophthongs, the correlation coefficients are within the range 0.63 to 0.88; averaging across all 12 monophthongs of each speaker results in a coefficient of 0.94. Furthermore, it appeared that the objective spectral analysis is as reliable as a subjective assessment by magnitude estimation by two to three listeners. Using principal components analysis (PCA), the number of dimensions by which the vowel spectra are described can be reduced. For the various monophthongs the range of the correlation coefficients between subjective distances and objective distances in a two-dimensional PCA subspace is 0.30-0.93. The three groups of speakers can still be distinguished in this subspace. In the extreme case of the deaf speakers all vowels are strongly "neutralized," whereas the different vowels of the native speakers are well separated, especially after speaker normalization; results are less clear for the non-natives.