初學小提琴常見演奏錯誤之自動辨識

Abstract

分析並建立演奏錯誤的模型，是以電腦輔助樂器學習不可或缺的一部份。本論文針對小提琴之初學者，呈現一套能夠自動辨識四種常見演奏錯誤的系統。我們首先建立一個包含980個小提琴音檔的資料集，其中每一個音檔都是一個以圓滑奏(legato)演奏的音符，並且由十位不同程度的演奏者錄製而成。我們讓小提琴演奏專家針對每一個音符聆聽、並標記可能的演奏錯誤。我們自同一組特徵集(feature set)萃取五種特徵表示的方法(feature representation)用以測試演奏錯誤之自動辨識系統，包含兩種音符層級(note-level)和三種片段層級(segment-level)：起始(onset)、持續(sustain)和結束(offset)的表示方法。我們的實驗架構分別以Fisher score以及支持向量機(support vector machine)作為特徵選取(feature selection)和分類演算法，並且以F-measure作為評估效能的數值。實驗結果顯示不同的特徵表示方法能夠於兩種演奏錯誤呈現近似17.70%的差異，反映出特定的特徵表示方法與演奏錯誤之關聯。再者，我們個別針對演奏錯誤探討其相關之特徵，結果顯示泛用的特徵如梅爾倒頻譜係數(MFCCs)並不足以解決我們的研究問題，因此需要進階的特徵設計或學習方法。最後，我們的演算法能夠優於演奏者的標記至30.79%，表示利用本實驗所建立之模型有益於小提琴之練習。

Analyzing and modeling playing mistakes are essential parts of computer-aided education tools in learning musical instruments. In this thesis, we present a system for identifying four types of mistakes commonly made by novice violin players. We construct a new dataset comprising of 980 legato notes played by 10 players across different skill levels, and have violin experts annotate all possible mistakes associated with each note by listening to the recordings. Five feature representations are generated from the same feature set with different scales, including two note-level representations and three segment-level representations of the onset, sustain and offset, and are tested for automatically identifying playing mistakes. Experiments are carried under the framework of using Fisher score for feature selection and the support vector machine for classification. Performance is evaluated in terms of F-measure. Results show that different feature representations can vary up to 17.70% for two types of playing mistakes. It demonstrates the different sensitivities of each feature representation to different mistakes. Moreover, we investigate the relevant features of each playing mistake and suggest that the standard audio features such as MFCCs are not good enough and more advanced feature design may be needed. Last but not least, the algorithm can outperform the players' self-annotations by 30.79%. It suggests the advantage of utilizing the model in violin practicing.