哼唱检索中旋律特征的聚类与优化方法

Melody Feature Clustering and Optimization for Query-by-humming

哼唱检索是音频检索的一个重要分支,其能够为用户提供一种方便快捷的全新体验。在检索过程中,由于同首歌的不同哼唱版本之间具有不容忽视的差异,因此对旋律特征进行精确匹配并无法得到理想的检索结果。针对这一问题,将基于优化初始聚类中心的k-means(optimized initial clustering center k-means,OICC k-means)聚类方法引入到哼唱检索系统中,通过对旋律特征进行聚类来充分学习不同旋律特征之间的结构相似性,从而将具有相似结构的旋律特征划分到同一聚类内给聚类编号,以为后端的旋律特征匹配提供更有效的标签。同时,考虑到聚类后的旋律特征可以进行进一步的特征表示,因此将聚类后的标签作为深度置信网络(deep belief networks,DBN)的输入标签并进行特征提取,以获取具有更强区分性的高层旋律特征,从而有效提升旋律特征的鲁棒性。在获取高层旋律特征后,需将聚类类别作为匹配标签,并进行哼唱检索即可。实验结果表明所提出的方法能够有效提升哼唱检索系统的性能。

Query-by-humming is an important branch of audio retrieval,and it can provide users with a new and convenient experience.During the retrieval process,since there are unignorable differences between different humming of the same song,it is difficult to obtain ideal results by accurate matching of the melody features.To solve this problem,the optimized initial clustering center k-means(OICC k-means)is introduced in the query-by-humming system.By clustering melody features to fully learn the structural similarity between different melody features,the melody features with similar structures are divided into the same cluster and the clusters are numbered,so as to provide more effective matching of melody features at the back end label.Meanwhile,considering that the clustered melody features can be further characterized,the clustered tags are used as the input tags of deep belief networks(DBN)and feature extraction is performed to obtain stronger discriminative characteristics the high-level melody features of the melody,thereby effectively improving the robustness of melody features.After obtaining the high-level melody features,it is necessary to use the cluster category as a matching tag and perform a humming search.Experimental results show that the proposed method can effectively improve the performance of humming retrieval system.