基于凝聚信息瓶颈的音频事件聚类方法

摘要

In order to further improve the performance of methods for audio events clustering,this paper proposes a method for audio events clustering based on the theory of agglomerative information bottleneck.First,the principles and derivations of information bottleneck are briefly introduced.Then,the proposed method is described in detail,including the definitions of source variables,relevance variables and destination variables,the steps of the proposed method and the analyses of main computational loads of all methods.The proposed method and two kinds of previous methods (including the method based on spectral clustering,and the method based on both Bayesian information criterion and agglomerative hierarchical clustering) are evaluated on the experimental data extracted from two different corpora of audio events.The experimental results show that the proposed method obtains higher K values (geometric mean of average clustering purity and average audio purity) and runs faster than the previous methods under several experimental conditions.%为了进一步提高音频事件聚类算法性能,本文基于凝聚信息瓶颈理论提出一种音频事件聚类方法.首先,论述信息瓶颈原理及其推导过程;然后,详细论述一种基于凝聚信息瓶颈的音频事件聚类方法,包括源变量、相关变量和目标变量的定义,聚类的具体步骤,算法主要计算量分析等.采用取自两个数据库的音频事件样本进行测试,实验结果表明:与目前文献报道的方法相比,本文方法在多种实验条件下都获得了更高的K值(平均类纯度和平均音频纯度的几何平均值),而且运算速度更快.