| 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 15:15 ［ポスター講演］Unsupervised Phoneme Segmentation Using Mahalanobis Distance ○Yu Qiao・Nobuaki Minematsu（Univ. of Tokyo）  SP2007-198 |
|
| 抄録 | （和） | One of the fundamental problems in speech engineering is phoneme segmentation. Approaches to phoneme segmentation can be divided into two categories: supervised and unsupervised segmentation. The approach of this paper belongs to the 2nd category: that is we try to perform phonetic segmentation without using any prior knowledge on linguistic contents and acoustic models. In an earlier work, we have formulated the segmentation problem into a probabilistic optimization problem by using statistics and information theory analysis. We developed an objective function: summation of square error (SSE) based on Euclidean distance of cepstrual features. However, it is not known whether or not Euclidean distance yields the best distance metric to estimate the goodness of the segments. A popular generalization of Euclidean distance is Mahalanobis distance. In this paper, we study whether and how Mahalanobis distance can be used to improve the performance of segmentation. The essential problem here is how to determine the parameters (covariance matrix) for Mahalanobis distance calculation. We deal with this problem in a learning based framework and develop two criteria for determining the optimal parameters: MSV and MDV. MSV minimizes the summation of variance within-phoneme, and MDV tries to maximize the ratio of the variance between phonemes to the variance within phonemes. Both of them can lead to close form solutions by using matrix calculation. We carried out experiments on the TIMIT database to compare the proposed methods. The results indicate that the using of learning Mahalanobis distance can improve the segmentation performance. |
| （英） | One of the fundamental problems in speech engineering is phoneme segmentation. Approaches to phoneme segmentation can be divided into two categories: supervised and unsupervised segmentation. The approach of this paper belongs to the 2nd category: that is we try to perform phonetic segmentation without using any prior knowledge on linguistic contents and acoustic models. In an earlier work, we have formulated the segmentation problem into a probabilistic optimization problem by using statistics and information theory analysis. We developed an objective function: summation of square error (SSE) based on Euclidean distance of cepstrual features. However, it is not known whether or not Euclidean distance yields the best distance metric to estimate the goodness of the segments. A popular generalization of Euclidean distance is Mahalanobis distance. In this paper, we study whether and how Mahalanobis distance can be used to improve the performance of segmentation. The essential problem here is how to determine the parameters (covariance matrix) for Mahalanobis distance calculation. We deal with this problem in a learning based framework and develop two criteria for determining the optimal parameters: MSV and MDV. MSV minimizes the summation of variance within-phoneme, and MDV tries to maximize the ratio of the variance between phonemes to the variance within phonemes. Both of them can lead to close form solutions by using matrix calculation. We carried out experiments on the TIMIT database to compare the proposed methods. The results indicate that the using of learning Mahalanobis distance can improve the segmentation performance. | |
| キーワード | （和） | Unsupervised phoneme segmentation / Optimization / Mahalanobis distance / Learning distance metric / / / / |
| （英） | Unsupervised phoneme segmentation / Optimization / Mahalanobis distance / Learning distance metric / / / / | |
| 文献情報 | 信学技報, vol. 107, no. 551, SP2007-198, pp. 69-74, 2008年3月. | |
| 資料番号 | SP2007-198 | |
| 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 | SP2007-198 |
|
| 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） | Unsupervised Phoneme Segmentation Using Mahalanobis Distance |
| サブタイトル（英） | |
| キーワード(1)（和/英） | Unsupervised phoneme segmentation / Unsupervised phoneme segmentation |
| キーワード(2)（和/英） | Optimization / Optimization |
| キーワード(3)（和/英） | Mahalanobis distance / Mahalanobis distance |
| キーワード(4)（和/英） | Learning distance metric / Learning distance metric |
| 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） | Yu Qiao / Yu Qiao / |
| 第1著者 所属（和/英） | the University of Tokyo (略称： 東大) the University of Tokyo (略称： Univ. of Tokyo) |
| 第2著者 氏名（和/英/ヨミ） | Nobuaki Minematsu / Nobuaki Minematsu / |
| 第2著者 所属（和/英） | the University of Tokyo (略称： 東大) the University of Tokyo (略称： Univ. of Tokyo) |
| 講演者 | 第1著者 |
| 発表日時 | 2008-03-20 15:15:00 |
| 発表時間 | 90分 |
| 申込先研究会 | SP |
| 資料番号 | SP2007-198 |
| 巻番号（vol） | vol.107 |
| 号番号（no） | no.551 |
| ページ範囲 | pp.69-74 |
| 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]