Title: 基於機器學習的自動拍號判定系統
Automatic Meter-Finding System Based on Machine Learning
Authors: 陳恒劭
Chen, Heng-Shao
Chang, Wen-Whei
Keywords: 樂器數位介面;自相關係數;區域邊界偵測模型;樂句特徵;支持向量機;深層類神經網路;MIDI;autocorrelation function;local boundary detection model;phrase-related feature;support vector machine;deep neural networks
Issue Date: 2017
Abstract: 拍號是演奏者對於樂曲理解與詮釋的重要依據,但大多數MIDI音樂並沒有記錄拍號資訊,因而造成音樂表現評量及內涵式資訊檢索的諸多限制。因此本論文針對電子鋼琴音樂,開發一套以MIDI音樂為基礎的自動拍號判定系統。有別於前人研究聚焦於不同加權的自相關係數,我們導入新的樂句特徵參數,有效改善2/4與4/4兩種拍號易於混淆而判定效果不佳之瓶頸。首先使用區域邊界偵測模型計算各個音符的邊界強度,接著根據2/4與4/4拍在樂句延伸度與小節架構上的差異性,選取適當代表點的邊界強度組成其樂句特徵參數。在系統製作上,我們結合樂句特徵與加權自相關係數,並使用支持向量機與類神經網路建構的拍號判定模型。根據200首四種拍號樂曲的實驗結果顯示,結合樂句特徵及二層隱藏層類神經網路的判定模型取得最高的91.20%正確率,其中2/4與4/4拍的判定正確率分別提升了86%與20%。
Meter is an important basis for pianists to understand and interpret music performance. However, the meter information is usually not available in MIDI music, limiting its applications in performance evaluation and context-based information retrieval. Goal of this study is to develop an automatic meter-finding system. Compared to previous works that focus on various accent types of autocorrelation function, we introduce the phrase-related features based on the music theory. Firstly, we apply the local boundary detection model to compute the boundary strength of each note. Then, we select the representative points according to different extension of the phrases and structural differences of bars. Finally, the boundary strengths corresponding to representative points are used to form the phrase-related features. Together with the autocorrelation function, we apply the support vector machine and deep neural networks to construct the meter-finding model. Experiment results show that the combined use of phrase-related features and autocorrelation function can improve the accuracy for 2/4 and 4/4 music by 86% and 20%, respectively. Also, the use of 2-hidden layers neural network achieves the highest accuracy rate of 91.20% among four different meters.
URI: http://etd.lib.nctu.edu.tw/cdrfb3/record/nctu/#GT070450745
Appears in Collections:Thesis