Abstract: Content-based music retrieval gains more and more significance in recent years，and a number of retrieval methods have been proposed.Most existing methods use only one or two significant features to represent a piece of music.Different from the existing methods，in this paper，a new method is proposed that represents a piece of music by a feature matrix generated from its spectrogram，thus the similarity between the query and any candidate music file in the database is evaluated by the similarity of two matrices.Experiments show that this method can achieve satisfactory retrieval performance.

Key words: acoustic music, spectrogram, matrix similarity, content-based music retrieval

摘要： 近年来，基于内容的音乐检索研究日益受到重视，不少检索方法被提出来。其中，大部分方法主要集中在精确地表征音乐的某一两个特征，以反映出音乐某一两个突出方面的性质。论文采取完全不同的思路，使用从声谱图中提取的特征矩阵来表示音乐，查询片断与数据库中候选乐曲的相似度从而转化成两个特征矩阵间的相似度。实验结果表明：该方法不仅过程与计算简单，而且能够取得良好的检索效果。

关键词: 波形音乐, 声谱图, 矩阵相似度, 基于内容的音乐检索