基于MDCT域特征的MP3音乐分类

音乐分类是将用户输入的音乐信号与音乐库中的音乐文件进行匹配,找出相应的类别.传统的MP3分类的研究大多先把MP3文件解压缩到PCM文件,然后在PCM文件上进行特征提取,这种方法存在的问题是处理速度比较慢.本文中借鉴语音识别技术,提出了基于MDCT域的MP3音乐特征片段提取方法,然后利用MDCT域上的音乐片段特点表示MP3音乐特征属性,最后采用适应性较强的学习分类器对已经提取的音乐特征向量进行分类.实验先通过对4个歌手100首歌的学习,然后对未知20首歌进行分类,识别演唱歌手平均准确度达80%.

A MDCT Domain Feature-Based Approach in Classifying MP3 Song

Music classification is the procedure of matching user input music signal with a music library to identify the user music's corresponding category.Traditional MP3 Classification methods decode MP3 files to PCM file,which is the base of feature extraction.The main problem of these methods is their slow speed.The paper focuses on the slow problem and proposes its fast solution.Using the speech recognition technology,the method extracts music clips feature based on the MDCT domain,and then use the above features as the classification attribute.Finally,it adopts the strong adaptability LCS classifier to extract the music eigenvector for classification.The experiment firstly learn 100 songs of 4 singers,25 songs per each,and then classify 20 unknown singer songs.The average accuracy of identifying singer is over 80%.