Abstract: A new indexing method for content-based music information retrieval is presented, and it is implemented on a query by humming (QBH) system. The method makes use of pitch and rhythm to describe a piece of melody. Considering the in-coordination of pitch and inaccuracy of rhythm information during melody input, average pitch difference and average duration percentage difference are counted as M-tree indexing keys. Since many errors occur during inputting, a searching radius is dynamically defined for each query. Experiments show that the indexing structure saves searching time of original QBH system by approximately 60% for long query. At the same time, the success rate of matching still remains at about 80%.

Key words: content-based music information retrieval, query-by-humming, M-tree indexing

摘要： 本文提出一种新型的音乐信息检索索引方法，并将其应用于哼唱检索系统上，获得可喜的成果。该方法根据音乐旋律的特点，使用音高和音长描述一段乐曲。考虑到各种音乐输入方式存在的音高不一致性和节奏伸缩性，系统根据乐曲旋律特征的总体变化，将平均音高差和平均音长变化百分比作为关键字建立M树索引结构。由于各类输入旋律常常存在误差，本文定义一定的搜索半径，查找半径范围内与输入的音乐片断最相似的音乐文件。实验结果表明，该结构可以使系统提高60%以上的检索速度，并且搜索结果的正确率最高仍然可达80%。

关键词: 基于内容的音乐信息检索, 哼唱检索, M树索引