Vowel Classification using a Dispersion-Analysis Method
K. P. Li, A. S. House, George W. Hughes · The Journal of the Acoustical Society of America · 1968
The classification of vocalic data by automatic means is under study. The techniques of dispersion analysis are used to determine the set of principal components from data and then to investigate the distribution of spectral data on the principal components. The results given by the analysis are given linguistic interpretations, and a simple method is proposed to construct a decision-tree structure for classifying vowel spectra. The sequence of decisions is described and the results are compared linguistic descriptions involving Jakobsonian distinctive features. The spectral data under study are the rectified outputs of 35 bandpass filters covering the range 250 Hz to 10 kHz, and sampled 60 times/sec. The speech samples were all contributed by one talker. Approximately 1500 time samples were selected from the quasistationary portions of 10 common vowels of American English; each vowel occurred in 23 different utterances, and the phonetic context differed in each utterance. [This work was supported in part by the Air Force Cambridge Research Laboratories under contract.]