Carol Y. Espy-Wilson
Elec. Comput. and Systems Eng. Dept., Boston Univ., Boston, MA 02215
This study investigates the automatic classification of labial and alveolar stop consonants. The algorithm implemented was proposed by Lahiri et al. [J. Acoust. Soc. Am. 76, 391--404 (1984)], and uses relative and dynamic measures relating the spectral shape at the burst with that of the onset of the following vowel. The purpose of the measure is to capture that (1) for labials, the difference in energy between the burst release spectrum and the vowel spectrum is either less at low frequencies than at high frequencies or the difference is compatible in both frequency ranges and (2) for alveolars, the difference in energy between the burst release spectrum and the vowel onset spectrum is less at high frequencies than at low frequencies. To make these distinctions, the algorithm draws slope lines on the burst and the vowel onset spectra based on F2 and F4 and computes a ratio of the difference (in dB) of the slope lines at 3500 and 1500 Hz. In the automatic procedure, F2 and F4 are picked from LPC spectra based on the covariance method. Preliminary results obtained from word-initial stops consonants occurring in monosyllabic words show that the algorithm does classify correctly many of the stop consonants. However, an analysis of the errors suggests the need for refinements. In particular, the alveolars classified as labials had less energy than the vowel between 0 and 4000 Hz (falling slope line), but more energy than the vowels at frequencies higher than 4000 Hz. In the case of labials, the energy difference at low frequencies was not less than that at high frequencies, but, the labials did have comparable or less energy to that of the vowel at frequencies above 4000 Hz. The refinements suggested by these findings will be made and results will be presented.