{"links":{},"metadata":{"_oai":{"id":"oai:ipsj.ixsq.nii.ac.jp:00236025","sets":["6504:11678:11689"]},"path":["11689"],"owner":"44499","recid":"236025","title":["実時間で動作する音響イベント検出の大規模事前学習"],"pubdate":{"attribute_name":"公開日","attribute_value":"2024-03-01"},"_buckets":{"deposit":"1032bcba-8830-4cd4-aeac-7474fdc96825"},"_deposit":{"id":"236025","pid":{"type":"depid","value":"236025","revision_id":0},"owners":[44499],"status":"published","created_by":44499},"item_title":"実時間で動作する音響イベント検出の大規模事前学習","author_link":["645021","645022","645020","645023"],"item_titles":{"attribute_name":"タイトル","attribute_value_mlt":[{"subitem_title":"実時間で動作する音響イベント検出の大規模事前学習"}]},"item_keyword":{"attribute_name":"キーワード","attribute_value_mlt":[{"subitem_subject":"人工知能と認知科学","subitem_subject_scheme":"Other"}]},"item_type_id":"22","publish_date":"2024-03-01","item_language":{"attribute_name":"言語","attribute_value_mlt":[{"subitem_language":"jpn"}]},"item_22_text_3":{"attribute_name":"著者所属","attribute_value_mlt":[{"subitem_text_value":"筑波大"},{"subitem_text_value":"産総研"},{"subitem_text_value":"同志社大，産総研"},{"subitem_text_value":"産総研"}]},"item_publisher":{"attribute_name":"出版者","attribute_value_mlt":[{"subitem_publisher":"情報処理学会","subitem_publisher_language":"ja"}]},"publish_status":"0","weko_shared_id":-1,"item_file_price":{"attribute_name":"Billing file","attribute_type":"file","attribute_value_mlt":[{"url":{"url":"https://ipsj.ixsq.nii.ac.jp/record/236025/files/IPSJ-Z86-1R-06.pdf","label":"IPSJ-Z86-1R-06.pdf"},"date":[{"dateType":"Available","dateValue":"2024-07-03"}],"format":"application/pdf","filename":"IPSJ-Z86-1R-06.pdf","filesize":[{"value":"442.0 kB"}],"mimetype":"application/pdf","accessrole":"open_date","version_id":"dd552707-bf6e-4429-b316-258fc69e7ecc","displaytype":"detail","licensetype":"license_note","license_note":"Copyright (c) 2024 by the Information Processing Society of Japan"}]},"item_22_creator_5":{"attribute_name":"著者名","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"大田, 竹蔵"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"坂東, 宜昭"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"井本, 桂右"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"大西, 正輝"}],"nameIdentifiers":[{}]}]},"item_resource_type":{"attribute_name":"資源タイプ","attribute_value_mlt":[{"resourceuri":"http://purl.org/coar/resource_type/c_5794","resourcetype":"conference paper"}]},"item_22_source_id_9":{"attribute_name":"書誌レコードID","attribute_value_mlt":[{"subitem_source_identifier":"AN00349328","subitem_source_identifier_type":"NCID"}]},"item_22_description_7":{"attribute_name":"論文抄録","attribute_value_mlt":[{"subitem_description":"本稿では実時間で動作する音響イベント検出 (SED) について述べる．SEDでは Audio Spectrogram Transformer (AST) が高い性能を発揮しているが，ASTは観測信号全体を入力するオフライン型であり，実時間推論で必要な逐次処理には不向きだった．そこで本研究では，入力を短時間のチャンクに分割する逐次型ASTによるSEDを構築する．具体的には，チャンク内の情報を次チャンクへ伝播させる記憶トークンを導入し，逐次処理でも高い性能を維持する．さらに，学習済みのオフライン型ASTの知識蒸留および1万時間を超える大規模学習データでの分散学習を併用して性能を改善した．","subitem_description_type":"Other"}]},"item_22_biblio_info_10":{"attribute_name":"書誌情報","attribute_value_mlt":[{"bibliographicPageEnd":"366","bibliographic_titles":[{"bibliographic_title":"第86回全国大会講演論文集"}],"bibliographicPageStart":"365","bibliographicIssueDates":{"bibliographicIssueDate":"2024-03-01","bibliographicIssueDateType":"Issued"},"bibliographicIssueNumber":"1","bibliographicVolumeNumber":"2024"}]},"relation_version_is_last":true,"weko_creator_id":"44499"},"updated":"2025-01-19T09:25:29.223597+00:00","created":"2025-01-19T01:37:51.209681+00:00","id":236025}