文献类型：期刊文献

中文题名：面向翻唱歌曲识别的相似度融合算法

英文题名：Similarity Distance Fusion Algorithm in Cover Song Identification

作者：刘婷[1];陈宁[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237

年份：2016

卷号：42

期号：6

起止页码：845

中文期刊名：华东理工大学学报（自然科学版）

外文期刊名：Journal of East China University of Science and Technology

收录：CSTPCD;;Scopus;北大核心:【北大核心2014】；CSCD:【CSCD2015_2016】；

基金：国家自然科学基金(61271349)

语种：中文

中文关键词：相似度融合;节拍追踪;瞬时频率音级轮廓;和声音级轮廓;耳蜗音级轮廓;Qmax;翻唱歌曲识别

外文关键词：similarity distance fusion; beat tracing;IF-PCP; HPCP; CPCP; Qmax; cover song identification

摘要：提出了一种面向翻唱歌曲识别的相似度融合算法。该算法将基于乐理特征的相似度和基于人耳感知特性的相似度融合,通过把基于节拍跟踪和瞬时频率音级轮廓(IF-PCP)的最大互相关相似度、基于和声音级轮廓(HPCP)的Qmax相似度、基于耳蜗音级轮廓(CPCP)的Q_(max)相似度映射到同一个多维空间,并计算其几何距离来进行相似度融合。该算法使得IF-PCP特征的节拍速度不变性、HPCP特征的和声优势、CPCP特征的人耳感知特性有效融合。为了验证算法的有效性,采用包含212首不同歌曲共502个版本的数据库作为测试对象,以平均正确率均值和TOP-N作为测试指标对算法性能进行测试。测试结果表明,与基于单一相似度算法相比,该融合算法可提高翻唱歌曲识别准确率。
This paper proposes a new similarity distance fusion algorithm that fuses the similarity distance of music theory feature and auditory perceptual feature. In the proposed algorithm, three similarity distances, IF-PCP based on beat tracing with maximum cross-correlation measure, HPCP with Qmax measure,and CPCP with Qmax measure, are projected in a multi-dimensional space and then the geometric distance as the fusion similarity distance is computed. This algorithm can effectively integrate the beat speed invariance of IF-PCP, the harmonic advantage of HPCP, and the auditory perceptual of CPCP. An experiment on a database with 502 versions of 212 different songs is made in this work. By mean of MAP and TOP-N as the performance indicator of the cover song identification, it is shown that the proposed algorithm in this paper can improve the precision of cover song identification greatly.