講演抄録/キーワード |
講演名 |
2008-03-20 15:15
[ポスター講演]LVCSR based on Context-Dependent Syllable Acoustic Models ○Jian Zhang・Longbiao Wang・Seiichi Nakagawa(Toyohashi Univ. of Tech.) SP2007-200 |
抄録 |
(和) |
We propose an effective and accurate inter-word context dependent modeling for large vocabulary continuous speech recognition (LVCSR). As well known, intra-word context dependent modeling can be realized by describing the context dependent syllables in the dictionary. However, it usually suffers from the limitations of less accuracy because it does not model inter-syllable pronunciation variations. In our previous study, a combinational use of linear lexicon and tree-structured lexicon in a 1-best approximation search algorithm for LVCSR was proposed. We only need to make branches for the head syllable according to the contexts and the paths are merged at the second syllable for the linear lexicon. For the tree-structured lexicon, branches are made in a similar way. At the end node of a word the language scores have to be compensated considering the inter-word context, but the scores of contexts other than that of the best history have lost because of the merge at the second syllable. To solve this problem, we introduce the `likelihood difference index'. We also investigate the effect of rescoring of the context dependent syllable acoustic model in the 2nd pass mode. The proposed algorithms were evaluated on JNAS and CSJ corpus. The proposed algorithms obtained a remarkable improvement of recognition performance, and the rescoring of the context dependent syllable acoustic models in the 2nd pass mode also achieved a further improvement even the same acoustic models were used in the 1st pass. |
(英) |
We propose an effective and accurate inter-word context dependent modeling for large vocabulary continuous speech recognition (LVCSR). As well known, intra-word context dependent modeling can be realized by describing the context dependent syllables in the dictionary. However, it usually suffers from the limitations of less accuracy because it does not model inter-syllable pronunciation variations. In our previous study, a combinational use of linear lexicon and tree-structured lexicon in a 1-best approximation search algorithm for LVCSR was proposed. We only need to make branches for the head syllable according to the contexts and the paths are merged at the second syllable for the linear lexicon. For the tree-structured lexicon, branches are made in a similar way. At the end node of a word the language scores have to be compensated considering the inter-word context, but the scores of contexts other than that of the best history have lost because of the merge at the second syllable. To solve this problem, we introduce the `likelihood difference index'. We also investigate the effect of rescoring of the context dependent syllable acoustic model in the 2nd pass mode. The proposed algorithms were evaluated on JNAS and CSJ corpus. The proposed algorithms obtained a remarkable improvement of recognition performance, and the rescoring of the context dependent syllable acoustic models in the 2nd pass mode also achieved a further improvement even the same acoustic models were used in the 1st pass. |
キーワード |
(和) |
LVCSR / context-dependent acoustic models / acoustic and language model rescoring / linear lexicon / tree-structured lexicon / / / |
(英) |
LVCSR / context-dependent acoustic models / acoustic and language model rescoring / linear lexicon / tree-structured lexicon / / / |
文献情報 |
信学技報, vol. 107, no. 551, SP2007-200, pp. 81-86, 2008年3月. |
資料番号 |
SP2007-200 |
発行日 |
2008-03-13 (SP) |
ISSN |
Print edition: ISSN 0913-5685 Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
PDFダウンロード |
SP2007-200 |
|