話者適応における学習データ内の音素数と認識精度の考察

摘要

This study pay attention to number of phones in training sets. We make a hybrid HMM which mixed speaker adaptation HMM and independent HMM by number of phones. In addition, we make training sets by given bias to number of phones to obtain the high performance in speaker adaptation. As a result of experiments, we have improvement of 2.65% and 1.33% for 164 words training sets and 82 words training sets in speaker adaptation using a hybrid HMM and made training sets.%本研究は，学習データに含まれる音素数に着目し，音素数により話者適応HMMと不特定話者HMMを組み合わせた混合HMMを作成する．また，話者適応の認識精度がより高くなるように，各音素の数に偏りを持たせることで学習データを作成する．実験の結果，作成した混合HMMと偏りを持たせた学習データを用いて，164単語の話者適応において2．65％，82単語の話者適応において1．33％の改善が得られた．