| 電子情報通信学会 研究会発表申込システム 講演論文 詳細 |
技報閲覧サービス [ログイン] 技報アーカイブ |
| トップに戻る | 前のページに戻る | [Japanese] / [English] |
| 講演抄録／キーワード | ||
| 講演名 | 2008-03-20 13:00 ［招待講演］Feature Vector Classification based on Likelihood Ratio Jeong-sik Park・Sang-min Yoon・○Yung-Hwan Oh（KAIST）  SP2007-187 |
|
| 抄録 | （和） | This paper proposes a new feature vector classification method to reduce the recognition error due to overlap data in speaker or emotion recognition. The recognition process is composed of two distinct phases, training and test. During training phase, feature vectors are extracted from each speaker and emotion data and the respective speaker or emotion model (GMM) is constructed. In test phase, feature vector extracted from test utterance is applied to each model, and a model which has the maximum likelihood is decided as a result. The ideal training is constructing each model, minimizing overlap amongst models owing to several factors such as speaker or emotion similarities. Thus overlap disturbs classifying the GMMs and may degrade recognition accuracy. To reduce the effect of overlap, we classified feature vectors from training data into two categories, non-overlapped and overlapped, based on likelihood ratio. By using these separated feature vectors, we constructed a non-overlapped and an overlapped model, respectively, for each speaker and emotion. In test phase, we selected meaningful data from test utterance depending on likelihood ratio calculated from two kinds of models and used for decision-making. This method has an effect discarding overlap between models and thus improving recognition accuracy. We confirmed the efficiency of our method from speaker and emotion recognition experiments. |
| （英） | This paper proposes a new feature vector classification method to reduce the recognition error due to overlap data in speaker or emotion recognition. The recognition process is composed of two distinct phases, training and test. During training phase, feature vectors are extracted from each speaker and emotion data and the respective speaker or emotion model (GMM) is constructed. In test phase, feature vector extracted from test utterance is applied to each model, and a model which has the maximum likelihood is decided as a result. The ideal training is constructing each model, minimizing overlap amongst models owing to several factors such as speaker or emotion similarities. Thus overlap disturbs classifying the GMMs and may degrade recognition accuracy. To reduce the effect of overlap, we classified feature vectors from training data into two categories, non-overlapped and overlapped, based on likelihood ratio. By using these separated feature vectors, we constructed a non-overlapped and an overlapped model, respectively, for each speaker and emotion. In test phase, we selected meaningful data from test utterance depending on likelihood ratio calculated from two kinds of models and used for decision-making. This method has an effect discarding overlap between models and thus improving recognition accuracy. We confirmed the efficiency of our method from speaker and emotion recognition experiments. | |
| キーワード | （和） | / / / / / / / |
| （英） | / / / / / / / | |
| 文献情報 | 信学技報, vol. 107, no. 551, SP2007-187, pp. 5-10, 2008年3月. | |
| 資料番号 | SP2007-187 | |
| 発行日 | 2008-03-13 (SP) | |
| ISSN | Print edition: ISSN 0913-5685 Online edition: ISSN 2432-6380 | |
| 著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します．(許諾番号：10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) | |
| PDFダウンロード | SP2007-187 |
|
| 研究会情報 | |
| 研究会 | SP |
| 開催期間 | 2008-03-20 - 2008-03-21 |
| 開催地（和） | 東大 |
| 開催地（英） | Univ. Tokyo |
| テーマ（和） | 国際ワークショップ(20日），研究会(21日) 「聴覚・音声・言語とその障害，一般」 |
| テーマ（英） | International Workshop (Mar 20), Speech Production, Speech Perception, Hearing and Speech, etc. (Mar 21) |
| 講演論文情報の詳細 | |
| 申込み研究会 | SP |
| 会議コード | 2008-03-SP |
| 本文の言語 | 英語 |
| タイトル（和） | |
| サブタイトル（和） | |
| タイトル（英） | Feature Vector Classification based on Likelihood Ratio |
| サブタイトル（英） | |
| キーワード(1)（和/英） | / |
| キーワード(2)（和/英） | / |
| キーワード(3)（和/英） | / |
| キーワード(4)（和/英） | / |
| キーワード(5)（和/英） | / |
| キーワード(6)（和/英） | / |
| キーワード(7)（和/英） | / |
| キーワード(8)（和/英） | / |
| 第1著者 氏名（和/英/ヨミ） | Jeong-sik Park / Jeong-sik Park / |
| 第1著者 所属（和/英） | Korea Advanced Institute of Science and Technology (略称： KAIST) Korea Advanced Institute of Science and Technology (略称： KAIST) |
| 第2著者 氏名（和/英/ヨミ） | Sang-min Yoon / Sang-min Yoon / |
| 第2著者 所属（和/英） | Korea Advanced Institute of Science and Technology (略称： KAIST) Korea Advanced Institute of Science and Technology (略称： KAIST) |
| 第3著者 氏名（和/英/ヨミ） | Yung-Hwan Oh / Yung-Hwan Oh / |
| 第3著者 所属（和/英） | Korea Advanced Institute of Science and Technology (略称： KAIST) Korea Advanced Institute of Science and Technology (略称： KAIST) |
| 講演者 | 第3著者 |
| 発表日時 | 2008-03-20 13:00:00 |
| 発表時間 | 60分 |
| 申込先研究会 | SP |
| 資料番号 | SP2007-187 |
| 巻番号（vol） | vol.107 |
| 号番号（no） | no.551 |
| ページ範囲 | pp.5-10 |
| ページ数 | 6 |
| 発行日 | 2008-03-13 (SP) |
[研究会発表申込システムのトップページに戻る]