摘要
分析基于内容的音乐信息检索(music information retrieval,MIR),其关键在于特征提取.传统的单特征向量表示方法存在局限性:难以选定用于提取特征的片段或时间窗;只选取音乐片段会丢失一些重要的信息.为了消除局限性,引入多特征向量的特征表示方法,在获取音乐的多个声学特征向量的同时,也可以完整地表示该音乐曲目.为了更加准确地计算由多特征向量表示的2个音乐曲目之间的相似度,引入金字塔匹配核技术(pyramid match kernel,PMK)计算不同长度的多特征向量之间的相似度.实验结果表明,PMK技术的引入可以提高MIR的性能.
One key step in the content-based music information retrieval （MIR） is feature extraction. The limitations of the tradition- al feature representation-single-feature-vector are obvious. First, it is difficult to select the segment/time window for extracting acous- tic features second,some important music information may be lost due to the selection of the specified segment. To address these two limitations,a novel feature representation method-multi-feature-vector in which multiple acoustic feature vectors over the music track can be obtained,was introduced to represent the music tracks. Motivated by the pyramid match kernel （PMK） which measures the similarity between two multi-feature-vectors with different sizes,PMK was used to extend the MIR techniques, so that the similarity between two music tracks which are represented by nmlti-feature vectors can be accurately captured. Empirical results show that the incorporation of PMK into MIR tasks leads to the performance improvement.
出处
《厦门大学学报（自然科学版）》
CAS
CSCD
北大核心
2013年第1期32-37,共6页
Journal of Xiamen University：Natural Science
基金
福建省自然科学基金项目(2011J05157)
关键词
多特征向量
特征相似度
金字塔匹配核
音乐信息检索
multi-feature vector
feature similarity
pyramid match kernel
music information retrieval
