Investigating neural source-filter waveform model for statistical parametric speech synthesis

• Item type: SIG Technical Reports(1)
• 公開日: 2019-02-20
• タイトル: Investigating neural source-filter waveform model for statistical parametric speech synthesis
• タイトル / 言語: en
• タイトル: Investigating neural source-filter waveform model for statistical parametric speech synthesis
• 言語: eng
• 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 Institute of Informatics
• 著者所属: National Institute of Informatics
• 著者所属: National Institute of Informatics
• 著者所属(英): en
• 著者所属(英): National Institute of Informatics
• 著者所属(英): en
• 著者所属(英): National Institute of Informatics
• 著者所属(英): en
• 著者所属(英): National Institute of Informatics
• 著者名: Xin, Wang; Shinji, Takaki; Junichi, Yamagishi
• 著者名(英): Xin, Wang; Shinji, Takaki; Junichi, Yamagishi
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Recently we proposed the neural source-filter model (NSF) that converts a sequence of acoustic features into a speech waveform. Similar to other recent neural waveform models, the NSF is a non-autogressive model powered by dilated CNN; however, the NSF uses the sine waveform instead of the random noise as the excitation. Furthermore, without using the normalizing flow, the NSF simply optimizes the network parameters by minimizing a spectral amplitude distance. In this work, we further investigated the three issues: whether the network structure can be further simplified; whether the NSF can be applied to multi-speaker speech synthesis; whether the NSF can be directly applied to convert the linguistic features into the speech waveforms. Our experiments showed positive results on all the three points. Particularly, we found that the WaveNet-style gated activation can be safely removed, and the NSF performs quite well as a pure dilated-CONV-based network.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Recently we proposed the neural source-filter model (NSF) that converts a sequence of acoustic features into a speech waveform. Similar to other recent neural waveform models, the NSF is a non-autogressive model powered by dilated CNN; however, the NSF uses the sine waveform instead of the random noise as the excitation. Furthermore, without using the normalizing flow, the NSF simply optimizes the network parameters by minimizing a spectral amplitude distance. In this work, we further investigated the three issues: whether the network structure can be further simplified; whether the NSF can be applied to multi-speaker speech synthesis; whether the NSF can be directly applied to convert the linguistic features into the speech waveforms. Our experiments showed positive results on all the three points. Particularly, we found that the WaveNet-style gated activation can be safely removed, and the NSF performs quite well as a pure dilated-CONV-based network.
• 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 2019-SLP-126, 号 3, p. 1-5, 発行日 2019-02-20
• 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