電子鋼琴演奏熟練度的自動評分系統

Abstract

由於電子音樂科技的快速發展，透過網路分享MIDI演奏音樂資料的需求日益劇增。為了幫助初學者養成正確演奏方式與降低學習難度，鋼琴彈奏技巧熟練度的自動評分系統是一個相當重要的研究議題。本論文針對數位樂器介面MIDI格式的電子鋼琴音樂，利用MIDI特徵紀錄值進行三次雲規曲線擬合處理以取得演奏趨勢的特徵參數，接著分別利用決策樹與K近鄰這兩種機器學習演算法進行熟練度評分。以音階為輸入的實驗結果顯示，K近鄰評分機制與專家平均評分的相關係數為0.6543，決策樹評分機制則進一步提升其相關係數為0.7391。這兩組評分結果與不同專家間的評分呈現高度相關，由此可知本系統在MIDI音樂熟練度評分具有良好的一致性與可靠性。

In recent years, the low-priced MIDI piano has become one of the most popular musical instruments. However, expert pianists are not usually available when the piano beginners are in need of corrective actions for proficiency improvement. To construct a support system for self-learning at the beginning stage, we proposed an automatic proficiency evaluation for a scale performance within one octave by the MIDI-piano. We began by extracting five MIDI-based parameters concerning on-set time, velocity, duration, legato and tempo. Based on their deviations from constant standards, we apply the cubic spline curve fitting technique to extract the feature parameters representing the tendencies of the proficiency performance. Two machine learning algorithms, K-nearest neighbor (KNN) and decision tree, are employed to predict the proficiency scores of testing data. Their performance comparison was carried out in terms of the correlation coefficients between the evaluation scores given by automatic evaluation and four expert pianists. Experimental results on 11 subjects show that correlation coefficients obtained for KNN and decision tree are 0.6543 and 0.7391, respectively, indicating the effectiveness of the proposed methods.