基于频域卷积信号盲源分离的乐曲数据库构建  

Construction of song database based on frequency domain convolutive signal blind source separation

作　　者：李鹏[1] 周明全[2] 黎南杉[2] 王学松[1] 

机构地区：[1]北京师范大学虚拟现实与可视化技术研究所,北京100875 [2]北京师范大学信息科学与技术学院,北京100875

出　　处：《计算机应用研究》2010年第4期1376-1379,共4页Application Research of Computers

基　　金：国家自然科学基金面上资助项目(60673100);国家自然科学基金重点资助(60736008);北京市自然科学基金重点资助项目(4081002)

摘　　要：将通过频域卷积信号盲源分离算法从MP3歌曲音频信号中分离出人声主唱信号,再从人声主唱信号中提取出能够表征歌曲的旋律特征构建哼唱检索系统的歌曲数据库。盲源分离要求观测信号数目不小于源信号数目,因此先用小波多分辨率分析构造一路观测信号,再用频域独立成分分析(FDICA)实现MP3歌曲音频信号的盲源分离(BSS)。实验证明,采用FDICA-basedBSS从歌曲MP3中分离出的人声主唱信号的旋律特征与待检索的人声哼唱信号的旋律特征有较高的相似度,可以用歌曲MP3构建哼唱检索系统的歌曲数据库。This article separated the vocal melody signals from MP3 songs audio signals based on frequency domain convolutive signal blind source algorithm,and then built a song database of humming retrieval system using song melody features extracted from the vocal melody signal.Blind source separation requirements the number of observation signals was not less than the number of source signal,thus structured a observation signal by wavelet multi-resolution analysis at first,and then achieved the song＇s MP3 audio signal blind source separation through the frequency domain independent component analysis.Experiments show that the song melody characters extracted from the song vocal melody signal separated from the song MP3 by FDICA-based BSS are highly similar to the melody characters of retrieved humming signal,so the song database of humming retrieval system can be constructed using the song MP3.

关 键 词：卷积信号盲源分离 基于FDICA的盲源分离 小波多分辨率分析 歌曲数据库 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]