SBERT-based Musical Components Estimation from Lyrics Trained with Imbalanced “Orpheus” Data

• Item type: SIG Technical Reports(1)
• 公開日: 2023-06-16
• タイトル: SBERT-based Musical Components Estimation from Lyrics Trained with Imbalanced “Orpheus” Data
• タイトル / 言語: en
• タイトル: SBERT-based Musical Components Estimation from Lyrics Trained with Imbalanced “Orpheus” Data
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: 一般発表
• 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: Department of Computer and Network Engineering, The University of Electro-Communications
• 著者所属: Department of Computer and Network Engineering, The University of Electro-Communications
• 著者所属: Department of Data Science, Faculty of Business Administration, Asia University
• 著者所属: Department of Computer and Network Engineering, The University of Electro-Communications
• 著者所属: Department of Computer and Network Engineering, The University of Electro-Communications
• 著者所属(英): en
• 著者所属(英): Department of Computer and Network Engineering, The University of Electro-Communications
• 著者所属(英): en
• 著者所属(英): Department of Computer and Network Engineering, The University of Electro-Communications
• 著者所属(英): en
• 著者所属(英): Department of Data Science, Faculty of Business Administration, Asia University
• 著者所属(英): en
• 著者所属(英): Department of Computer and Network Engineering, The University of Electro-Communications
• 著者所属(英): en
• 著者所属(英): Department of Computer and Network Engineering, The University of Electro-Communications
• 著者名: Mastuti, Puspitasari; Takuya, Takahashi; Gen, Hori; Shigeki, Sagayama; Toru, Nakashika
• 著者名(英): Mastuti, Puspitasari; Takuya, Takahashi; Gen, Hori; Shigeki, Sagayama; Toru, Nakashika
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: This research was done to develop neural models that are capable of estimating appropriate musical components based on lyrics input. We extracted paired data of lyrics and musical components from “Orpheus”, a Japanese automated composition system with over 6000 user-published songs on the platform and used them as training data. These lyrics are converted into text embeddings with Sentence-BERT and then fed into neural models with their respective musical components for training. The imbalance in the data is mitigated by using focal loss to avoid overfitting and the performance of our models are evaluated subjectively through a survey. These models can be implemented in automated composition system to provide automated setup recommendation for the users and or used as a source of inspiration in conventional composition.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: This research was done to develop neural models that are capable of estimating appropriate musical components based on lyrics input. We extracted paired data of lyrics and musical components from “Orpheus”, a Japanese automated composition system with over 6000 user-published songs on the platform and used them as training data. These lyrics are converted into text embeddings with Sentence-BERT and then fed into neural models with their respective musical components for training. The imbalance in the data is mitigated by using focal loss to avoid overfitting and the performance of our models are evaluated subjectively through a survey. These models can be implemented in automated composition system to provide automated setup recommendation for the users and or used as a source of inspiration in conventional composition.
• 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: AN10438388
• 書誌情報: 研究報告音楽情報科学（MUS）; 巻 2023-MUS-137, 号 57, p. 1-5, 発行日 2023-06-16
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8752
• 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