一覧に戻る

タイトル
  • en An improved minimum generation error based model adaptation for HMM-based speech synthesis
作成者
    • en Wu, Yi-Jian ja Wu, Yi-Jian
    • en Qin, Long
内容注記
  • Other en Aminimum generation error (MGE) criterion had been proposedfor model training in HMM-based speech synthesis. Inthis paper, we apply the MGE criterion to model adaptation forHMM-based speech synthesis, and introduce an MGE linear regression(MGELR) based model adaptation algorithm, wherethe regression matrices used to transform source models are optimizedso as to minimize the generation errors of adaptationdata. In addition, we incorporate the recent improvements ofMGE criterion into MGELR-based model adaptation, includingstate alignment under MGE criterion and using a log spectraldistortion (LSD) instead of Euclidean distance for spectraldistortion measure. From the experimental results, the adaptationperformance was improved after incorporating these twotechniques, and the formal listening tests showed that the qualityand speaker similarity of synthesized speech after MGELRbasedadaptation were significantly improved over the originalMLLR-based adaptation.
  • Other en Brighton, United KingdomSeptember 6-10, 2009
出版者 en International Speech Communication Association
日付
    Issued2010-08-31
言語
  • eng
資源タイプ conference paper
出版タイプ VoR
資源識別子 URI https://nitech.repo.nii.ac.jp/records/3456
関連
  • URI http://www.isca-speech.org/archive/interspeech_2009/i09_1787.html http://www.isca-speech.org/archive/interspeech_2009/i09_1787.html
収録誌情報
  • en INTERSPEECH 2009 10th Annual Conference of the International Speech Communication Association
  • 開始ページ1787 終了ページ1790
ファイル
コンテンツ更新日時 2025-03-14