{"links":{},"id":184865,"metadata":{"_oai":{"id":"oai:ipsj.ixsq.nii.ac.jp:00184865","sets":["1164:5159:9063:9316"]},"path":["9316"],"owner":"11","recid":"184865","title":["CycleGANを用いた高品質なノンパラレル声質変換"],"pubdate":{"attribute_name":"公開日","attribute_value":"2017-12-14"},"_buckets":{"deposit":"b920e031-c027-46ad-b9dc-6f20c6fdcd35"},"_deposit":{"id":"184865","pid":{"type":"depid","value":"184865","revision_id":0},"owners":[11],"status":"published","created_by":11},"item_title":"CycleGANを用いた高品質なノンパラレル声質変換","author_link":["409480","409482","409477","409478","409481","409479"],"item_titles":{"attribute_name":"タイトル","attribute_value_mlt":[{"subitem_title":"CycleGANを用いた高品質なノンパラレル声質変換"},{"subitem_title":"High-quality nonparallel voice conversion using CycleGAN","subitem_title_language":"en"}]},"item_keyword":{"attribute_name":"キーワード","attribute_value_mlt":[{"subitem_subject":"ポスターセッション","subitem_subject_scheme":"Other"}]},"item_type_id":"4","publish_date":"2017-12-14","item_4_text_3":{"attribute_name":"著者所属","attribute_value_mlt":[{"subitem_text_value":"国立情報学研究所"},{"subitem_text_value":"国立情報学研究所／エジンバラ大学"},{"subitem_text_value":"国立情報学研究所"}]},"item_language":{"attribute_name":"言語","attribute_value_mlt":[{"subitem_language":"jpn"}]},"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/184865/files/IPSJ-SLP17119009.pdf","label":"IPSJ-SLP17119009.pdf"},"date":[{"dateType":"Available","dateValue":"2019-12-14"}],"format":"application/pdf","billing":["billing_file"],"filename":"IPSJ-SLP17119009.pdf","filesize":[{"value":"973.9 kB"}],"mimetype":"application/pdf","priceinfo":[{"tax":["include_tax"],"price":"660","billingrole":"5"},{"tax":["include_tax"],"price":"330","billingrole":"6"},{"tax":["include_tax"],"price":"0","billingrole":"22"},{"tax":["include_tax"],"price":"0","billingrole":"44"}],"accessrole":"open_date","version_id":"2c45434d-6b16-4d59-b1b0-c8226acbb4e5","displaytype":"detail","licensetype":"license_note","license_note":"Copyright (c) 2017 by the Information Processing Society of Japan"}]},"item_4_creator_5":{"attribute_name":"著者名","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"房, 福明"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"山岸, 順一"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"越前, 功"}],"nameIdentifiers":[{}]}]},"item_4_creator_6":{"attribute_name":"著者名(英)","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"Fuming, Fang","creatorNameLang":"en"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"Junichi, Yamagishi","creatorNameLang":"en"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"Isao, Echizen","creatorNameLang":"en"}],"nameIdentifiers":[{}]}]},"item_4_source_id_9":{"attribute_name":"書誌レコードID","attribute_value_mlt":[{"subitem_source_identifier":"AN10442647","subitem_source_identifier_type":"NCID"}]},"item_4_textarea_12":{"attribute_name":"Notice","attribute_value_mlt":[{"subitem_textarea_value":"SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc."}]},"item_resource_type":{"attribute_name":"資源タイプ","attribute_value_mlt":[{"resourceuri":"http://purl.org/coar/resource_type/c_18gh","resourcetype":"technical report"}]},"item_4_source_id_11":{"attribute_name":"ISSN","attribute_value_mlt":[{"subitem_source_identifier":"2188-8663","subitem_source_identifier_type":"ISSN"}]},"item_4_description_7":{"attribute_name":"論文抄録","attribute_value_mlt":[{"subitem_description":"近年，機械学習の進展により声質変換の性能が大幅に向上した．しかし，学習データが対とならないノンパラレルの場合，ソース話者とターゲット話者の特徴を精密にマッチすることが難しい．ノンパラレル声質変換モデルの学習はまだ困難であり，変換性能はまだ低い問題がある．一方，画像変換分野ではペアなしの画像データベースから変換モデルを学習する方法として CycleGAN が注目されている．CycleGAN は GAN の一種であり，複数個の generator と discriminator を持つ．また，generator は入力データの一部の情報を維持しながら，discriminator との競争学習によりターゲットドメインへの変換ができる特徴がある．そこで，本研究はこのアイディアに基づいて CycleGAN をノンパラレル声質変換に適用する方法を提案する．提案手法では，ソース話者とターゲット話者の類似特徴を直接マッチするのではなく，ソース話者の一部の言語情報を維持しながら話者特徴をターゲット話者にできるだけ近付けるように変換モデルを学習する．被験者評価実験より，提案手法は標準の GAN に基づいたパラレル声質変換を上回ったことを示す．","subitem_description_type":"Other"}]},"item_4_description_8":{"attribute_name":"論文抄録(英)","attribute_value_mlt":[{"subitem_description":"Recently, voice conversion (VC) based on deep learning has achieved remarkable performance. However, it is still difficult to train a mapping model using nonparallel training samples. In this work, we propose a high-quality nonparallel VC training method based on CycleGAN. A CycleGAN is a kind of generative adversarial network (GAN) originally developed for unpaired image-to-image translation. This model can be learned by an approach that a part of input information is kept while the corresponding distribution of the input data can be converted into a target distribution without paired training samples. Experimental results show that the proposed method outperforms a standard GAN-based parallel VC system.","subitem_description_type":"Other"}]},"item_4_biblio_info_10":{"attribute_name":"書誌情報","attribute_value_mlt":[{"bibliographicPageEnd":"6","bibliographic_titles":[{"bibliographic_title":"研究報告音声言語情報処理（SLP）"}],"bibliographicPageStart":"1","bibliographicIssueDates":{"bibliographicIssueDate":"2017-12-14","bibliographicIssueDateType":"Issued"},"bibliographicIssueNumber":"9","bibliographicVolumeNumber":"2017-SLP-119"}]},"relation_version_is_last":true,"weko_creator_id":"11"},"created":"2025-01-19T00:52:08.756232+00:00","updated":"2025-01-20T03:09:16.089664+00:00"}