摘要
针对哼唱的非精确性,提出一种新的方法,将模糊集合及方法引入旋律近似匹配的过程,在用户查询与数据库目标数据之间构造模糊隶属函数,根据隶属度判断音高差信息的相似度,同时对音长比信息进行相似度计算,用两者加权得分作为动态规划法匹配过程中的转换代价,最终得到两个匹配序列的编辑距离,从而得到查询结果.实验结果显示模糊方法的引入提高了检索命中率,同时考虑音长特征的策略也提高了检索精度.在样本集内,检索结果的前十位命中率较原有的5阶音高差近似匹配方法提高了16%.
Query-by-humming is a much friendlier method for content based music information retrieval systems. Since an input humming query may have various errors, the system should be able to tolerate these errors. This paper represents a new melody matching method for the QBH based MIR system. The fuzzy relationship is established between the input query and the target data in the database by constructing the membership function of the fuzzy subset to judge the similarity of the delta pitches to its target. Both delta pitch and pitch duration ratio are used in melody representation. The DP matching method is taken to do the similarity measurements. Experimental results show the effectiveness of the new method, its top-10 success rate exceeding the traditional 5-level contour approximate matching method by 16%.
出处
《西安电子科技大学学报》
EI
CAS
CSCD
北大核心
2006年第1期85-88,102,共5页
Journal of Xidian University
基金
"十五"国家部委科技电子预研基金资助项目(413160501)
关键词
哼唱检索
模糊方法
旋律表示
旋律匹配
query-by-humming
fuzzy approach
melody representation
melody matching
