本論文的主題是建構一個音樂檢索系統。主要的概念是透過MPEG-7音樂特徵值(Audio Descriptor)來判別某歌曲片段是否存在於歌曲資料庫當中。然而此系統能否達到實用的目標,端賴是否有高效率之搜尋方法。若歌曲片段比對花費太多時間,則會降低系統的實用性。我們以MPEG-7之聲音簽章特徵值(Audio Signature Descriptor)為基礎,提出降低資料維度的方法,並且利用多維度最近鄰居搜尋法(Multidimensional Nearest Neighbor Searching)加快搜尋速度,讓整體比對時間能大幅降低。我們也提出一改善誤警率(False Alarm Rate)的方法,進而降低錯誤接收率(FAR)與錯誤拒絕率(FRR),並以實驗探討其有效性。最後,我們利用多重解析度搜尋技巧實做本系統。
In this thesis, we propose a musical retrieval system. The main concept is to identify whether one piece of sound track is the same as another one in the song database by using MPEG-7 audio descriptor. However, the practicability of this system is based on whether it has some efficient searching method. If the comparison between query song and songs in database costs too much time, it will decrease system’s practicability. Based on Audio Signature Descriptor, we propose some methods about dimension reduction and the use of KD-tree for multidimensional nearest neighbor searching. It decreases the overall comparison time to increases practical value of our system. We also use some methods to improve system’s false alarm rate (i.e., decrease FAR and FRR) and benchmark those methods by ROC graph. Finally, we use multi-resolution search to implement our system.