多集団の線形判別関数による母音スペクトルの特徴抽出

概要

論文の詳細を見る
This paper describes a new method of feature extraction from speech spectra based on multi-category linear discriminant function. Five Japanese vowels pronounced by 24 male speakers and 26 female speakers were analyzed by a filter bank (Fig. 1). Then four components were extracted from the output of the filter bank by this method and a vowel feature space was constructed by these four components (Fig. 3). 96% of these vowels were identified correctly using only two of the four components and 98. 6% using all the four components. It was shown that this method gives a better result than principal-component analysis under the same condition (Figs. 4 and 5). The coefficients calculated from the sample vowels were also applicable to new sample vowels (Figs. 6 and 7). Furthermore, the first and the second components were found to be highly correlated with the first and the second formant frequencies, respectively (Figs. 8~11). The trajectories of combined vowels /ie/, /ia/, /io/ and /iu/ in the first and the second component plane were similar to their trajectories in the F_1 - F_2 diagram (Fig. 12). From these experimental results, this method is expected to be useful for real time feature extraction from speech signals.
社団法人日本音響学会の論文
1973-08-01