@techreport{oai:ipsj.ixsq.nii.ac.jp:00075438,
 author = {千々岩, 圭吾 and 鈴木, 雅之 and 齋藤, 大輔 and 峯松, 信明 and 広瀬, 啓吉 and Keigo, Chijiiwa and Masayuki, Suzuki and Daisuke, Saito and Nobuaki, Minematsu and Keikichi, Hirose},
 issue = {15},
 month = {Jul},
 note = {特徴量変換による雑音抑制手法である SPLICE は比較的少ない計算量で高い性能を発揮する.しかし,この SPLICE にも雑音の変動に頑健でないという問題がある.今回は変換方法そのものを未知の雑音環境に適応することを試みた.本報告では,主成分分析を用いて,未知の雑音環境下において推定すべきパラメータを削減し,少数の適応データで適応できる手法を提案する.また,AURORA-2 データベース testb セットにおいて提案手法を評価し,理想的な適応データが得られた場合には 12.0%,粗い雑音推定を用いた場合でも 1.9% の誤り削減率を得られた., SPLICE is one of the speech enhancement methods using feature conversion, which shows a high performance with a relationally small amount of calculation. It supposes that input noisy environments are static and similar to training data environments, so it is not guaranteed to work well in unseen environments. Therefore, we propose a new method to adapt conversion functions to work well in unseen environments. The proposed method reduces the number of parameters through Principal Component Analysis, so that the adapted conversion function is obtained with a smaller amount of adaptive data. Experiments show 12.0% reduction of error rate in the ideal adaptive condition and 1.9% reduction even in unfavorable conditions.},
 title = {Eigen-SPLICEを用いた雑音環境下における音声認識},
 year = {2011}
}