Using Functional Load for Optimizing DPGMM based Zero Resource Sub-word Unit Discovery

• Item type: SIG Technical Reports(1)
• 公開日: 2018-12-03
• タイトル: Using Functional Load for Optimizing DPGMM based Zero Resource Sub-word Unit Discovery
• タイトル / 言語: en
• タイトル: Using Functional Load for Optimizing DPGMM based Zero Resource Sub-word Unit Discovery
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: セッション2 単語獲得・感情認識
• 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: Nara Institute of Science and Technology
• 著者所属: Nara Institute of Science and Technology／RIKEN, Center for Advanced Intelligence Project AIP
• 著者所属: Beijing Language and Culture University
• 著者所属: Nara Institute of Science and Technology／RIKEN, Center for Advanced Intelligence Project AIP
• 著者所属(英): en
• 著者所属(英): Nara Institute of Science and Technology
• 著者所属(英): en
• 著者所属(英): Nara Institute of Science and Technology / RIKEN, Center for Advanced Intelligence Project AIP
• 著者所属(英): en
• 著者所属(英): Beijing Language and Culture University
• 著者所属(英): en
• 著者所属(英): Nara Institute of Science and Technology / RIKEN, Center for Advanced Intelligence Project AIP
• 著者名: Bin, Wu; Sakriani, Sakti; Jinsong, Zhang; Satoshi, Nakamura
• 著者名(英): Bin, Wu; Sakriani, Sakti; Jinsong, Zhang; Satoshi, Nakamura
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Unsupervised sub-word discovery of the zero resource language gains attention recently. One of the methods to tackle the problem is using an unsupervised clustering algorithm to recover the discrete phone-like units from the speech, such as the Dirichlet Process Gaussian Mixture Model (DPGMM), which currently achieves top results in the Zero Resource Speech Challenge. However, the DPGMM model is too sensitive to the acoustic variation and often produces too many types of sub-word units. This paper proposes to apply functional load to reduce the size of sub-word units from DPGMM. The functional load is the measurement of how much information in communication is conveyed by contrasts of these units. Then, the aim is to ignore the contrasts of the sub-word units that contribute little in conveying the information of the speech leading to decrease of the number of sub-word classes. We experiment on the official Zerospeech 2015 measuring with ABX error rate.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Unsupervised sub-word discovery of the zero resource language gains attention recently. One of the methods to tackle the problem is using an unsupervised clustering algorithm to recover the discrete phone-like units from the speech, such as the Dirichlet Process Gaussian Mixture Model (DPGMM), which currently achieves top results in the Zero Resource Speech Challenge. However, the DPGMM model is too sensitive to the acoustic variation and often produces too many types of sub-word units. This paper proposes to apply functional load to reduce the size of sub-word units from DPGMM. The functional load is the measurement of how much information in communication is conveyed by contrasts of these units. Then, the aim is to ignore the contrasts of the sub-word units that contribute little in conveying the information of the speech leading to decrease of the number of sub-word classes. We experiment on the official Zerospeech 2015 measuring with ABX error rate.
• 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 2018-SLP-125, 号 4, p. 1-2, 発行日 2018-12-03
• 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