・名古屋工業大学学術機関リポジトリは、名古屋工業大学内で生産された学術情報を電子的に収集・保存・発信するシステムです。 ・論文の著作権は、著者または出版社が保持しています。著作権法で定める権利制限規定を超える利用については、著作権者に許諾を得てください。 ・著者版フラグに「author」と記載された論文は、著者原稿となります。実際の出版社版とは、レイアウト、字句校正レベルの異同がある場合もあります。 ・Nagoya Institute of Technology Repository Sytem is built to collect, archive and offer electronically the academic information produced by Nagoya Institute of Technology. ・The copyright and related rights of the article are held by authors or publishers. The copyright owners' consents must be required to use it over the curtailment of copyrights. ・Textversion "Author " means the article is author's version. Author version may have some difference in layouts and wordings form publisher version.
INTERSPEECH 2009 10th Annual Conference of the International Speech Communication Association
ページ
680 - 683
発行年
2009
出版者
International Speech Communication Association
著者版フラグ
publisher
内容記述
This paper proposes a deterministic annealing based training algorithmfor Bayesian speech recognition. The Bayesian methodis a statistical technique for estimating reliable predictive distributionsby marginalizing model parameters. However, the localmaxima problem in the Bayesian method is more serious thanin the ML-based approach, because the Bayesian method treatsnot only state sequences but also model parameters as latentvariables. The deterministic annealing EM (DAEM) algorithmhas been proposed to improve the local maxima problem in theEM algorithm, and its effectiveness has been reported in HMMbasedspeech recognition using ML criterion. In this paper, theDAEM algorithm is applied to Bayesian speech recognition torelax the local maxima problem. Speech recognition experimentsshow that the proposed method achieved a higher performancethan the conventional methods.
Brighton, United KingdomSeptember 6-10, 2009