音樂分段在研究音樂分析相關的領域是很重要的研究題目。音樂的分段可以提供作音樂結構分析、音樂瀏覽、音樂內容查詢與音樂摘要等應用。本論文的研究目的就是對音樂作自動分段，以幫助使用者能快速瀏覽音樂的內容。因此，我們針對音樂的主題作主題式的分段。 音樂的主題是取決於作曲者的動機，動機是構成音樂主題的基本因素。為了能夠以主題為基礎作音樂分段，我們必須找出決定音樂主題的因素。動機會有規則性的出現在整首音樂當中，所以我們可以利用動機出現的規則來探勘音樂的動機。 我們提出一個以主題對音樂作分段的方法，總共分為四個主要的步驟。第一，我們從原始的音樂資料擷取出主旋律的部分。第二，將主旋律做粗略分段。我們利用探勘Non-trivial重複樣式的技術[17]，來找出粗略段落。第三，從粗略段落中探勘動機。我們利用Stein所提出來的動機變化規則，修改傳統探勘重複序列的方法，做動機的探勘。最後，我們利用探勘出來的動機對主旋律作精細分段。我們針對MIDI音樂檔案利用提出來的方法，實做出一個系統，找出音樂的主題段落。 先前研究在評估實驗結果時，多採用Precision與Recall去評估實驗的結果。然而，這樣的評估方法並不能表現出實驗結果與正確答案之間的相似程度。所以我們提出新的評估方法，根據實驗結果與正確答案之間的相似程度來評估實驗的準確率。根據實驗結果顯示，我們的方法準確率約65%。 / Music segmentation is one of the important issues in music analysis. Music segmentation can be utilized for music structure analysis, music browsing, content-based music retrieval, and music summarization. In this theis, we proposed a music segmentation method based on the music theme to provide users the capability to browse music segments by theme. Motives, the concepts of the composer, are the basic elements of music themes. Music themes were constructed by motives. In order to segment music by themes, we have to discover motives. Most motives repeated in the music by some motivic treatment rules. Therefore, motives can be discovered by these rules. We proposed the theme segmentation method. There are four steps. Firstly, we extract main melody from original music. In the second step, rough segments are generated from main melody by mining non-trivial repeating patterns. Then, motives are detected from rough segments. We modify the mining algorithm for discovering frequent patterns by applying motivic treatment rules proposed by Stein. Finally, we segment main melody based on the generated motives. Moreover, a system for segmentation of music in MIDI format was implemented. Concerning the effectiveness evaluation of music segmentation, precision and recall are used in previous research. We proposed an effectiveness measure and corresponding algorithm to evaluate the accuracy of music segmentation. Experimental results show that our proposed music segmentation method achieves 65% accuracy.