流行音樂之副歌偵測

Abstract

在流行音樂中，副歌通常代表整首歌中最精華且最容易記憶的部分。若我們可以自動擷取副歌的部分，當使用者在線上音樂軟體中搜尋音樂時，就可以透過試聽副歌的部分更快速的找到符合他們需求的音樂，或是可利用副歌的部分更進一步地分析歌曲的情緒，使得在音樂索取系統中可透過另一種途徑來管理或取得音樂的資訊。因此我們提出了一個在流行音樂中偵測副歌的系統，旨在擷取每段副歌，並可判斷副歌片段之數量。我們以前人的方法做為設計此系統之範本，利用由歌曲轉換而成的色彩圖像，將其分群，使得相似的音框可合為同一個片段，再把擁有平均相似度最高的一群判斷為副歌的部分，即可得數個副歌片段。再來，使用合併機制去除較細小的片段，以增加副歌片段之連續性及完整性。然後，精確每個副歌片段之起始及結束位置。最後，透過副歌片段彼此之間的匹配程度，捨棄偵測錯誤的副歌片段，並調整副歌片段之邊界位置，再取回遺漏的副歌片段。

On popular music, chorus sections are the most representative and memorable portions of a song. By automatically detecting chorus sections, when users search music in music browsers, they could pre-listen the chorus section and quickly find the desired song. Moreover, with chorus sections, we can progressively analyze the emotion of a song and increase a better approach to manage or access the music in music retrieval systems. Therefore, a chorus detection system on popular music is proposed and the purpose is not only to extract every chorus sections but also to detect the number of chorus sections. We utilize the method of the pioneer as our prototype. First, acquiring the colormap which is transformed from audio signal of popular music. Clustering the colormap in order to combine similar frames. Assign the cluster with largest similarity to be the chorus. Then the combination strategies are employed to get rid of tiny segments and increase the continuity of chorus sections. Refine the chorus sections to attain the accurate start points and end points. Finally, through the similarity between chorus sections, discarding the false sections, adjusting the boundary of chorus sections, and retrieve the missed chorus secionts.