基于模糊聚类的音乐哼唱检索的研究与实现

针对哼唱检索中匹配过程的耗时性和哼唱的非完全准确性,提出了并行模糊动态时间规整算法实现音乐的哼唱检索。采用相对音高差表示旋律,用动态时间规整算法实现哼唱音高差序列与目标音高差序列的近似匹配。在匹配过程中,引入模糊集合及模糊聚类,通过构造哼唱音高差与目标音高差之间的隶属函数并计算隶属度得到音高差信息的相似度,进而获得转换代价矩阵,最后得到两个匹配序列的匹配距离。为提高检索速度,引入并行算法实现匹配过程。实验结果表明,模糊方法的引入提高了检索精度,并行算法的运用明显缩短了检索时间。基于并行模糊动态时间规整算法的音乐哼唱检索的正确率达到72%左右,在双核计算机上进行实验,引入并行算法后检索时间缩短一半。

Research and Implementation of Music Humming Retrieval Based on Fuzzy Clustering

Aiming at the time-consuming matching process and non-completely accurate humming in music retrieval via Query By Humming(QBH),a parallel fuzzy Dynamic Time Warping(DTW) algorithm is proposed to implement approximate matching of melodies.Use relative pitch difference to represent melodies,and use DTW algorithm to realize approximate match between humming pitch difference sequences and target pitch difference sequences.In the process of matching,fuzzy sets and fuzzy clustering are introduced.By constructing membership functions between humming pitch difference and target pitch difference and computing membership grade,we get similarity degree of pitch differences,and then obtain the "Switching Cost" matrix.Finally,we get the matching distance of these two sequences.To accelerate the retrieval speed,parallel algorithms are introduced in the process of matching.The experiment results show that the introduction of fuzzy methods improves retrieval accuracy and the application of parallel algorithms shortens retrieval time.The accurate rate of humming retrieval based on parallel fuzzy DTW is about 72%,and in dual-core computer,the retrieval time is shorten a half after introducing parallel algorithm.