Recognition of vowels and a semivowel using formant locus extracted by cubic spline
M. Matsumura, Hiroaki Yamane, Katsuhiko Fujii · Electronics and Communications in Japan (Part III Fundamental Electronic Science) · 1989
Abstract The purpose of this study is to recognize the semivowel /j/ and vowels from the continuous speech of an unspecified speaker. the format frequencies are used as the feature parameters, which correspond to the articulation by the tongue and the lower jaw. In the extraction of the formant locus, the continuity is emphasized. A pre‐processing is made which detects the missing formant or unnecessary peaks of the speech spectrum. Then the interpolation and smoothing are made using the spline function. A triangle on the formant plane F1‐F2 plane) is considered, where the vertices represent the articulations of /a/, /i/ and /u/. the triangle is transformed into an equiliateral triangle. By this transformation, the speaker differences of the positions of the vowels on the F1‐F2 plane are normalized. the point minimizing the rate of change of the normalized formant frequency is determined, and, based on the result, the segmentation and recognition are performed. the recognition rate of 92.8 percent was achieved for the vowels in the continuous speech uttered by ten speakers. Then based on the analysis of articulation by tongue and lower jaw in pronouncing the semivowel /j/, a recognition experiment was attempted by extracting the feature parameters for discrimination from the shape of the formant locus. As a result, the semi‐vowel recognition rate of 94 percent was achieved for continuous speech uttered by nine speakers.