@techreport{oai:ipsj.ixsq.nii.ac.jp:00087740,
 author = {グェンドゥックズイ and 吉岡, 拓也 and 峯松, 信明 and 広瀬, 啓吉 and Duy, NguyenDuc and Takuya, Yoshioka and Nobuaki, Minematsu and Keikichi, Hirose},
 issue = {23},
 month = {Dec},
 note = {近年，音声認識技術は様々なアプリケーションで使用されている．しかし，録音環境に含まれる雑音や残響等の音響的な歪みにより認識性能が大幅に低下する．この問題の解決策として，クリーン音声の GMM を用いて観測音声の特徴量から音響的歪みの影響を取り除く特徴量強調技術が知られている．一方，モバイルデバイスへの音声入力に代表される最近のアプリケーションの多くでは，多様な環境で録られた認識対象個人の音声データを蓄積しておくことが容易にできる．しかしながら，こうした個人データをどのように扱えば特徴量強調を含む認識システム全体の性能を効果的に向上できるかは明らかでない．本研究では，特徴量強調に用いるクリーン音声 GMM の MAP 適応と音声認識に用いる音響モデルの MLLR 適応のいくつかの組み合わせ方について，その効果を実験的に比較検討する．, Speech recognition has been an active research area for many years, and nowadays, it is being used in many practical applications. However, the recognition performance is often seriously degraded by noise and reverberation present in recording environments. One promising approach to solve this problem is feature enhancement, which attempts to restore clean feature vectors using a GMM of clean speech. Meanwhile, in many recent applications including those to mobile devices, it is easy to collect target user's speech data recorded in various environments. However, how to exploit these data for improving the performance of a speech recognizer that performs feature enhancement is an open question. This study experimentally compares different methods for combining MAP adaptation of the clean speech GMM and MLLR adaptation of the recognizer's acoustic model.},
 title = {特徴量強調における教師なし話者適応に関する検討},
 year = {2012}
}