ログイン
言語:

WEKO3

  • トップ
  • ランキング
To
lat lon distance
To

Field does not validate



インデックスリンク

インデックスツリー

メールアドレスを入力してください。

WEKO

One fine body…

WEKO

One fine body…

アイテム

{"_buckets": {"deposit": "2c1e77f3-8f78-48d7-8303-06ec23165007"}, "_deposit": {"id": "12410", "owners": [], "pid": {"revision_id": 0, "type": "depid", "value": "12410"}, "status": "published"}, "_oai": {"id": "oai:nagoya.repo.nii.ac.jp:00012410", "sets": ["734"]}, "author_link": ["39070", "39071"], "item_12_alternative_title_19": {"attribute_name": "その他のタイトル", "attribute_value_mlt": [{"subitem_alternative_title": "音声認識における音響特徴変換の最適化基準の一般化に関する研究", "subitem_alternative_title_language": "ja"}]}, "item_12_biblio_info_6": {"attribute_name": "書誌情報", "attribute_value_mlt": [{"bibliographicIssueDates": {"bibliographicIssueDate": "2010-09-30", "bibliographicIssueDateType": "Issued"}}]}, "item_12_date_granted_64": {"attribute_name": "学位授与年月日", "attribute_value_mlt": [{"subitem_dategranted": "2010-09-30"}]}, "item_12_degree_grantor_62": {"attribute_name": "学位授与機関", "attribute_value_mlt": [{"subitem_degreegrantor": [{"subitem_degreegrantor_language": "ja", "subitem_degreegrantor_name": "名古屋大学"}, {"subitem_degreegrantor_language": "en", "subitem_degreegrantor_name": "Nagoya University"}], "subitem_degreegrantor_identifier": [{"subitem_degreegrantor_identifier_name": "13901", "subitem_degreegrantor_identifier_scheme": "kakenhi"}]}]}, "item_12_degree_name_61": {"attribute_name": "学位名", "attribute_value_mlt": [{"subitem_degreename": "博士(情報科学)", "subitem_degreename_language": "ja"}]}, "item_12_description_4": {"attribute_name": "抄録", "attribute_value_mlt": [{"subitem_description": "This thesis deals with acoustic feature transformations in automatic speech recognition to improve basic performance of a speech recognizer. The aim of acoustic feature transformations is to reduce dimensionality of long-term speech features without losing discriminative information among the different phonetic classes.\u003cbr/\u003eFirst, we focus on optimizing acoustic feature transformations using criteria with which to maximize the ratio of between-class scatter to within-class scatter. This approach is based on a family of functions of scatter or covariance matrices, which is frequently used in practice. Typical methods in this approach include linear discriminant analysis (LDA), heteroscedastic linear discriminant analysis (HLDA), and heteroscedastic discriminant analysis (HDA). Although LDA, HLDA and HDA are the most widely used in speech recognition, the connections between them have been disregarded so far. By developing a unified mathematical framework, close relationships between them are identified and analyzed in detail. The framework termed power LDA (PLDA) can describe various criteria by varying its control parameter. PLDA includes LDA, HLDA and HDA as special cases. In order to determine a sub-optimal control parameter automatically, a control parameter selection method is also provided.\u003cbr/\u003eThe effectiveness of the combinations of acoustic feature transformations and discriminative training techniques of acoustic models is investigated and additional performance improvement is obtained. Unfortunately, the transformation methods mentioned above may result in an unexpected dimensionality reduction if the data in a certain class consist of several clusters, because they implicitly assume that data are generated from a single Gaussian distribution. This study provides extensions of HDA and PLDA to deal with class distributions with several clusters.\u003cbr/\u003eSecond, we focus attention on acoustic feature transformations which minimize a kind of classification error between different phonetic classes. As the performance of speech recognition systems generally correlates strongly with the classification accuracy of features, the features should have the power to discriminate between different classes. The existing methods for this approach attempt to minimize the average classification error between different classes. Although minimizing the average classification error suppresses total classification error, it cannot prevent the occurrence of considerable overlaps between distributions of some different classes with low frequencies, which is critical for speech recognition because there may be class pairs that have little or no discriminative information on each other. Instead of the average classification error, minimization methods of maximum classification error are proposed herewith so as to avoid considerable error between different classes. In addition, interpolation methods that minimize the maximization error while minimizing the average classification error are also proposed and achieved the best results.", "subitem_description_language": "en", "subitem_description_type": "Abstract"}]}, "item_12_description_5": {"attribute_name": "内容記述", "attribute_value_mlt": [{"subitem_description": "名古屋大学博士学位論文 学位の種類:博士(情報科学)(課程) 学位授与年月日:平成22年9月30日", "subitem_description_language": "ja", "subitem_description_type": "Other"}]}, "item_12_dissertation_number_65": {"attribute_name": "学位授与番号", "attribute_value_mlt": [{"subitem_dissertationnumber": "甲第8967号"}]}, "item_12_identifier_60": {"attribute_name": "URI", "attribute_value_mlt": [{"subitem_identifier_type": "HDL", "subitem_identifier_uri": "http://hdl.handle.net/2237/14293"}]}, "item_12_select_15": {"attribute_name": "著者版フラグ", "attribute_value_mlt": [{"subitem_select_item": "publisher"}]}, "item_12_text_63": {"attribute_name": "学位授与年度", "attribute_value_mlt": [{"subitem_text_value": "2010"}]}, "item_access_right": {"attribute_name": "アクセス権", "attribute_value_mlt": [{"subitem_access_right": "open access", "subitem_access_right_uri": "http://purl.org/coar/access_right/c_abf2"}]}, "item_creator": {"attribute_name": "著者", "attribute_type": "creator", "attribute_value_mlt": [{"creatorNames": [{"creatorName": "坂井, 誠", "creatorNameLang": "ja"}], "nameIdentifiers": [{"nameIdentifier": "39070", "nameIdentifierScheme": "WEKO"}]}, {"creatorNames": [{"creatorName": "SAKAI, Makoto", "creatorNameLang": "en"}], "nameIdentifiers": [{"nameIdentifier": "39071", "nameIdentifierScheme": "WEKO"}]}]}, "item_files": {"attribute_name": "ファイル情報", "attribute_type": "file", "attribute_value_mlt": [{"accessrole": "open_date", "date": [{"dateType": "Available", "dateValue": "2018-02-20"}], "displaytype": "detail", "download_preview_message": "", "file_order": 0, "filename": "k8967.pdf", "filesize": [{"value": "839.7 kB"}], "format": "application/pdf", "future_date_message": "", "is_thumbnail": false, "licensetype": "license_note", "mimetype": "application/pdf", "size": 839700.0, "url": {"label": "k8967.pdf", "objectType": "fulltext", "url": "https://nagoya.repo.nii.ac.jp/record/12410/files/k8967.pdf"}, "version_id": "190da1e1-7248-42a1-a49f-50ed4b583f10"}]}, "item_language": {"attribute_name": "言語", "attribute_value_mlt": [{"subitem_language": "eng"}]}, "item_resource_type": {"attribute_name": "資源タイプ", "attribute_value_mlt": [{"resourcetype": "doctoral thesis", "resourceuri": "http://purl.org/coar/resource_type/c_db06"}]}, "item_title": "Acoustic Feature Transformation Based on Generalized Criteria for Speech Recognition", "item_titles": {"attribute_name": "タイトル", "attribute_value_mlt": [{"subitem_title": "Acoustic Feature Transformation Based on Generalized Criteria for Speech Recognition", "subitem_title_language": "en"}]}, "item_type_id": "12", "owner": "1", "path": ["734"], "permalink_uri": "http://hdl.handle.net/2237/14293", "pubdate": {"attribute_name": "PubDate", "attribute_value": "2010-10-27"}, "publish_date": "2010-10-27", "publish_status": "0", "recid": "12410", "relation": {}, "relation_version_is_last": true, "title": ["Acoustic Feature Transformation Based on Generalized Criteria for Speech Recognition"], "weko_shared_id": -1}
  1. A500 情報学部/情報学研究科・情報文化学部・情報科学研究科
  2. A500d 学位論文
  3. 博士論文(情科博・論情科博)

Acoustic Feature Transformation Based on Generalized Criteria for Speech Recognition

http://hdl.handle.net/2237/14293
http://hdl.handle.net/2237/14293
f7fd85ba-7bd2-4c97-ae7f-8136c1e2a149
名前 / ファイル ライセンス アクション
k8967.pdf k8967.pdf (839.7 kB)
Item type 学位論文 / Thesis or Dissertation(1)
公開日 2010-10-27
タイトル
タイトル Acoustic Feature Transformation Based on Generalized Criteria for Speech Recognition
言語 en
その他のタイトル
その他のタイトル 音声認識における音響特徴変換の最適化基準の一般化に関する研究
言語 ja
著者 坂井, 誠

× 坂井, 誠

WEKO 39070

ja 坂井, 誠

Search repository
SAKAI, Makoto

× SAKAI, Makoto

WEKO 39071

en SAKAI, Makoto

Search repository
アクセス権
アクセス権 open access
アクセス権URI http://purl.org/coar/access_right/c_abf2
抄録
内容記述 This thesis deals with acoustic feature transformations in automatic speech recognition to improve basic performance of a speech recognizer. The aim of acoustic feature transformations is to reduce dimensionality of long-term speech features without losing discriminative information among the different phonetic classes.<br/>First, we focus on optimizing acoustic feature transformations using criteria with which to maximize the ratio of between-class scatter to within-class scatter. This approach is based on a family of functions of scatter or covariance matrices, which is frequently used in practice. Typical methods in this approach include linear discriminant analysis (LDA), heteroscedastic linear discriminant analysis (HLDA), and heteroscedastic discriminant analysis (HDA). Although LDA, HLDA and HDA are the most widely used in speech recognition, the connections between them have been disregarded so far. By developing a unified mathematical framework, close relationships between them are identified and analyzed in detail. The framework termed power LDA (PLDA) can describe various criteria by varying its control parameter. PLDA includes LDA, HLDA and HDA as special cases. In order to determine a sub-optimal control parameter automatically, a control parameter selection method is also provided.<br/>The effectiveness of the combinations of acoustic feature transformations and discriminative training techniques of acoustic models is investigated and additional performance improvement is obtained. Unfortunately, the transformation methods mentioned above may result in an unexpected dimensionality reduction if the data in a certain class consist of several clusters, because they implicitly assume that data are generated from a single Gaussian distribution. This study provides extensions of HDA and PLDA to deal with class distributions with several clusters.<br/>Second, we focus attention on acoustic feature transformations which minimize a kind of classification error between different phonetic classes. As the performance of speech recognition systems generally correlates strongly with the classification accuracy of features, the features should have the power to discriminate between different classes. The existing methods for this approach attempt to minimize the average classification error between different classes. Although minimizing the average classification error suppresses total classification error, it cannot prevent the occurrence of considerable overlaps between distributions of some different classes with low frequencies, which is critical for speech recognition because there may be class pairs that have little or no discriminative information on each other. Instead of the average classification error, minimization methods of maximum classification error are proposed herewith so as to avoid considerable error between different classes. In addition, interpolation methods that minimize the maximization error while minimizing the average classification error are also proposed and achieved the best results.
言語 en
内容記述タイプ Abstract
内容記述
内容記述 名古屋大学博士学位論文 学位の種類:博士(情報科学)(課程) 学位授与年月日:平成22年9月30日
言語 ja
内容記述タイプ Other
言語
言語 eng
資源タイプ
資源 http://purl.org/coar/resource_type/c_db06
タイプ doctoral thesis
書誌情報
発行日 2010-09-30
学位名
言語 ja
学位名 博士(情報科学)
学位授与機関
学位授与機関識別子Scheme kakenhi
学位授与機関識別子 13901
言語 ja
学位授与機関名 名古屋大学
言語 en
学位授与機関名 Nagoya University
学位授与年度
学位授与年度 2010
学位授与年月日
学位授与年月日 2010-09-30
学位授与番号
学位授与番号 甲第8967号
著者版フラグ
値 publisher
URI
識別子 http://hdl.handle.net/2237/14293
識別子タイプ HDL
戻る
0
views
See details
Views

Versions

Ver.1 2021-03-01 18:53:58.668740
Show All versions

Share

Mendeley Twitter Facebook Print Addthis

Cite as

エクスポート

OAI-PMH
  • OAI-PMH JPCOAR
  • OAI-PMH DublinCore
  • OAI-PMH DDI
Other Formats
  • JSON
  • BIBTEX

Confirm


Powered by WEKO3


Powered by WEKO3