Discriminative Data Selection from Multiple ASR Systems' Hypotheses for Unsupervised Acoustic Model Training

• Item type: SIG Technical Reports(1)
• 公開日: 2015-11-25
• タイトル: Discriminative Data Selection from Multiple ASR Systems' Hypotheses for Unsupervised Acoustic Model Training
• タイトル / 言語: en
• タイトル: Discriminative Data Selection from Multiple ASR Systems' Hypotheses for Unsupervised Acoustic Model Training
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: 音声認識
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: School of Informatics, Kyoto University
• 著者所属: School of Informatics, Kyoto University
• 著者所属: School of Informatics, Kyoto University
• 著者所属(英): en
• 著者所属(英): School of Informatics, Kyoto University
• 著者所属(英): en
• 著者所属(英): School of Informatics, Kyoto University
• 著者所属(英): en
• 著者所属(英): School of Informatics, Kyoto University
• 著者名: Sheng, Li; Yuya, Akita; Tatsuya, Kawahara
• 著者名(英): Sheng, Li; Yuya, Akita; Tatsuya, Kawahara
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: This paper addresses unsupervised training of DNN acoustic model, by exploiting a large amount of unlabeled data with CRF-based classifiers. In the proposed scheme, we obtain ASR hypotheses by complementary GMM and DNN based ASR systems. Then, a set of dedicated classifiers are designed and trained to select the better hypothesis and verify the selected data. It is demonstrated that the classifiers can effectively filter usable data from unlabeled data for acoustic model training. The proposed method achieved significant improvement in the ASR accuracy from the baseline system, and it outperformed the models trained from the data selected based on the confidence measure scores (CMS) and also from the simple ROVER-based system combination.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: This paper addresses unsupervised training of DNN acoustic model, by exploiting a large amount of unlabeled data with CRF-based classifiers. In the proposed scheme, we obtain ASR hypotheses by complementary GMM and DNN based ASR systems. Then, a set of dedicated classifiers are designed and trained to select the better hypothesis and verify the selected data. It is demonstrated that the classifiers can effectively filter usable data from unlabeled data for acoustic model training. The proposed method achieved significant improvement in the ASR accuracy from the baseline system, and it outperformed the models trained from the data selected based on the confidence measure scores (CMS) and also from the simple ROVER-based system combination.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AN10442647
• 書誌情報: 研究報告音声言語情報処理（SLP）; 巻 2015-SLP-109, 号 8, p. 1-6, 発行日 2015-11-25
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8663
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会