Recent Advances of Exemplar-Based Sparse Representation for Voice Conversion

• Item type: SIG Technical Reports(1)
• 公開日: 2017-12-14
• タイトル: Recent Advances of Exemplar-Based Sparse Representation for Voice Conversion
• タイトル / 言語: en
• タイトル: Recent Advances of Exemplar-Based Sparse Representation for Voice Conversion
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: オーガナイズドセッション: 音声言語情報処理が切り拓く新しい情報社会
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: National University of Singapore
• 著者所属(英): en
• 著者所属(英): National University of Singapore
• 著者名: Haizhou, Li
• 著者名(英): Haizhou, Li
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: The exemplar-based sparse representation marks a success by performing voice conversion with a very limited amount of training data. In this framework, the activation matrix is shared between the source and target speakers. In a sense, such activation matrix serves as the bridge between the speakers. Ideally the activation matrix should capture the speaker invariant phonetic / prosodic information. Recently, phonetic and prosodic information has been introduced in different ways to the exemplar-based sparse representation framework to improve the voice conversion quality. In this talk, we will discuss the idea of speaker dependent and speaker independent features in voice conversion. We aim to carry over the speaker independent features from source to target, and transform the speaker dependent features of source to those of target. We will report the latest advances in phonetic sparse representation, and the use of Phonetic PosteriorGrams features and prosodic features, as the case studies on how to improve the activation matrix.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: The exemplar-based sparse representation marks a success by performing voice conversion with a very limited amount of training data. In this framework, the activation matrix is shared between the source and target speakers. In a sense, such activation matrix serves as the bridge between the speakers. Ideally the activation matrix should capture the speaker invariant phonetic / prosodic information. Recently, phonetic and prosodic information has been introduced in different ways to the exemplar-based sparse representation framework to improve the voice conversion quality. In this talk, we will discuss the idea of speaker dependent and speaker independent features in voice conversion. We aim to carry over the speaker independent features from source to target, and transform the speaker dependent features of source to those of target. We will report the latest advances in phonetic sparse representation, and the use of Phonetic PosteriorGrams features and prosodic features, as the case studies on how to improve the activation matrix.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AN10442647
• 書誌情報: 研究報告音声言語情報処理（SLP）; 巻 2017-SLP-119, 号 16, p. 1-1, 発行日 2017-12-14
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8663
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会