通常の164単語の学習データを用いた場合の実験結果を表 16に,偏りを持つ164単語の学習データを用いた場合の 実験結果を表17に示す.
10個未満 | 20個未満 | 30個未満 | 40個未満 | 50個未満 | ||
話者適応 | 混合HMM | 混合HMM | 混合HMM | 混合HMM | 混合HMM | |
mau | 86.87% | 90.69% | 91.45% | 91.45% | 91.30% | 91.18% |
(2276/2620) | (2376/2620) | (1396/2620) | (2396/2620) | (2392/2620) | (2389/2620) | |
mmy | 87.63% | 89.54% | 90.95% | 90.95% | 89.85% | 90.15% |
(2296/2620) | (2346/2620) | (2383/2620) | (2383/2620) | (2354/2620) | (2362/2620) | |
mnm | 85.50% | 88.97% | 89.20% | 89.20% | 87.63% | 88.70% |
(2240/2620) | (2331/2620) | (2337/2620) | (2337/2620) | (2296/2620) | (2324/2620) | |
faf | 87.98% | 91.30% | 91.95% | 91.95% | 90.80% | 90.73% |
(2305/2620) | (2392/2620) | (2409/2620) | (2409/2620) | (2379/2620) | (2377/2620) | |
fms | 84.77% | 89.73% | 90.61% | 90.61% | 89.39% | 90.34% |
(2221/2620) | (2351/2620) | (2374/2620) | (2374/2620) | (2342/2620) | (2367/2620) | |
ftk | 86.37% | 89.85% | 92.79% | 92.79% | 91.60% | 91.64% |
(2263/2620) | (2354/2620) | (2431/2620) | (2431/2620) | (2400/2620) | (2401/2620) | |
平均 | 86.52% | 90.01% | 91.16% | 91.16% | 90.10% | 90.46% |
(13601/15720) | (14150/15720) | (14330/15720) | (14330/15720) | (14163/15720) | (14220/15720) | |
10個未満 | 20個未満 | 30個未満 | 40個未満 | 50個未満 | ||
話者適応 | 混合HMM | 混合HMM | 混合HMM | 混合HMM | 混合HMM | |
mau | 91.64% | 91.64% | 92.06% | 92.06% | 91.56% | 91.26% |
(2401/2620) | (2401/2620) | (2412/2620) | (2412/2620) | (2399/2620) | (2391/2620) | |
mmy | 91.30% | 91.30% | 91.11% | 91.11% | 89.85% | 87.82% |
(2392/2620) | (2392/2620) | (2387/2620) | (2387/2620) | (2354/2620) | (2301/2620) | |
mnm | 89.85% | 89.85% | 90.15% | 90.15% | 88.85% | 87.75% |
(2354/2620) | (2354/2620) | (2362/2620) | (2362/2620) | (2328/2620) | (2299/2620) | |
faf | 91.72% | 91.72% | 92.60% | 92.60% | 91.91% | 90.11% |
(2403/2620) | (2403/2620) | (2426/2620) | (2426/2620) | (2408/2620) | (2361/2620) | |
fms | 89.05% | 89.05% | 89.54% | 89.54% | 88.40% | 87.98% |
(2333/2620) | (2333/2620) | (2346/2620) | (2346/2620) | (2316/2620) | (2305/2620) | |
ftk | 91.60% | 91.60% | 93.40% | 93.40% | 92.90% | 91.30% |
(2400/2620) | (2400/2620) | (2447/2620) | (2447/2620) | (2434/2620) | (2392/2620) | |
平均 | 90.86% | 90.86% | 91.48% | 91.48% | 90.58% | 89.37% |
(14283/15720) | (14283/15720) | (14380/15720) | (14380/15720) | (14239/15720) | (14049/15720) | |
164単語の混合HMMを用いた場合の6話者の平均誤り率を図14に, 偏りを持つ164単語の混合HMMを用いた場合の6話者の平均誤り率を図 15に示す.
結果より,30個未満混合HMMが最も高い認識精度だとわかる.40個未満・50個未 満混合HMMを用いた場合,30個未満混合HMMより認識精度が低下し不特定話者の認 識精度に近付いている.
本研究の条件では30個未満混合HMM(=30)を用いることで最も高い認識精度が 得られるが,このの最適値は,用いるデータベースや学習方法によって変化す ると考えらる.また音素の種類によっても,学習データ量による認識精度の違い があるため,音素ごとのの値を調べる必要があると考えている.