Analysis and recognition of Korean vowels
Bu-Il Kim, Hiroya Fujisaki · The Journal of the Acoustical Society of America · 1974
As a preliminary study to the automatic recognition of spoken Korean, the eight Korean vowels, uttered in isolation by 20 native speakers, including adults and youths of both sex, are analyzed and recognized. The speech materials are sampled at 10 kHz with 11-bit accuracy and the most stationary segment of 50-msec duration is selected from each utterance and converted into the logarithmic power spectrum. Frequencies of the first four formants are then accurately determined by the method of “analysis-by-synthesis,” using a model containing five formants and a correction factor for higher formants, as well as variable factors representing source and radiation characteristics. These vowel data are then classified by an optimum decision tree using 28 linear decision functions, each determined for the optimum separation of a pair of vowels in the three-dimensional space of the first three formant frequencies, giving a recognition score of 93%.