動的計画法を利用した音声の時間正規化に基づく連続単語認識

概要

論文の詳細を見る
We consider that the pattern matching method with time-normalizing ability is one of the most effective methods for the spoken word recognition, and that it can well be applied to the recognition of continuously spoken words. Speech can be expressed as a vector-valued time function (1), (2) by appropriate feature extraction. Then, the effect of speaking rate variation can be regarded as non-linear transformation of time axis (3), and can well be compensated by minimizing (4)(Fig. 1). Based on these consideration, we evaluate time-normalized similarity S(A, B) by (5). Calculation of S(A, B) is efficiently carried out using dynamic programming technique (15), (16), (17), (Fig. 3). Utilizing this pattern matching scheme, continuously spoken words can by separated into word units by determining the subpattern A^l which is most similar to the stored reference pattern, or for which S(A^l, B) is minimum (19), . (20), (Fig. 4). Based on this segmentation scheme, three methods to recognize continuously spoken words are proposed. Method-a is direct application of this segmentation scheme. In Metod-b, recognition is carried out by evaluating similarity S(A^l, B_mB_n) between concatenated reference patterns (17) and input pattern, and the amount of computation is considerably reduced using the segmentation scheme and the DP technique(Fig. 5). Method-c is modification of Method-b, where, after recognizing each word, matching window is shifted so that it will be able to cover the timing difference of the next word(Fig. 6). These Methods were extensively examined by computer simulation. Average recognition rate of 99. 8 per cent has been obtained for 2400 utterances of Japanese 1-digit numbers of five speakers, and 99. 6 per cent for total of 500 continuously spoken 2-digit numbers of five male speakers by Method-c(Table 1, 2, 3).
社団法人日本音響学会の論文
1971-09-10

動的計画法を利用した音声の時間正規化に基づく連続単語認識

スポンサーリンク

概要

著者

関連論文

スポンサーリンク