偏自己相関係数による数字音声認識の改良

概要

論文の詳細を見る
According to the result of our previous report, the partial autocorrelation coefficients (PAC) seem to be the best parameter for Japanese spoken digit (limited vocabulary) recognition. In this paper, further improvement of the recognition score and several methods to prevent the recognition score from deterioration, when the simplified method of PAC extraction is used, are described. We adopted the nonuniform time pattern matching algorithm. and speaker's own utterances were used as the standards. The extraction error of PAC increased when the speech wave amplitude is small (Fig. 1). So the contribution of the matching of PAC to the recognition decision is weighted by the form of w(t)=(1+a(t))/2, where a(t) represents the normalized relative amplitude (a(t)≦1). Relatively good result was obtained (Fig. 2). We checked some techniques for the prevention of score from deteriorating when fixed point arithmetics with a short accumulator length are used in the PAC extraction. The use of only reliable PAC in the recognition improve the recognition score remarkably, for example, 88% to 99% using 18 bits fixed point accumulator. Smoothing of parameters along the time axis is also effective (Fig. 3). A few other method are tried but the results are rather negative. After all, the two methods proved to be useful . . . 1) weighted matching of PAC by signal amplitude of standard utterance when computation accuracy of PAC extraction is adequately high, 2) A combination of the use of only reliable PAC and smoothing of each parameters in the time axis when the accuracy is relatively low. We verified the validity of these two methods through recognition experiments with various standard utterances of the same speaker (Fig. 7, 8, 9), different speakers (Fig. 8), and another extraction calculation method (Fig. 9).
社団法人日本音響学会の論文
1973-07-01

偏自己相関係数による数字音声認識の改良

スポンサーリンク

概要

著者

関連論文

スポンサーリンク