一种结合ctm和hmm的音乐分类方法amusic-bowenpublishing.pdfVIP

September 2012, Volume 2, Issue 3, PP.101-106 Scientific Journal of Computer Science 一种结合 CTM 和 HMM 的音乐分类方法 徐桂彬，邓伟 苏州大学 计算机科学与技术学院，江苏 苏州 215006 摘 要：研究一种相关主题模型 (correlated topic model, CTM) 与隐马尔可夫模型（hidden markov model, HMM）相结合的音 乐分类方法。这一方法首先将音乐片段的声学特征向量通过相似性比较转化为单词序列，然后通过CTM 将单词序列转换为主 题向量。最后，将这一主题向量输入HMM 得出分类结果。在这一方法中，用DBSCAN 聚类来确定CTM 的主题数。由于CTM 的建模，既大幅度降低了数据的维数，又保留了音乐片段的特性，因此不仅提高了方法的效率，而且增强了方法的分类能力。 实验验证了方法的有效性。 关键词: 音乐分类；相关主题模型；隐马尔可夫模型；DBSCAN 聚类 A Music Classification Method Combining CTM and HMM * Guibin Xu, Wei Deng School of Computer Science Technology, Soochow University, Suzhou, China, 215006 Email: dengw@ Abstract: In this paper, a new music classification method combining Correlated Topic Model (CTM) and Hidden Markov Model (HMM) is proposed. First, the acoustic feature vectors of a music clip are transformed into a sequence of words through replacing each feature vector with its most similar word in the vocabulary. Then, the sequence of words is fed into CTM model to infer a low-dimensional topic vector of the music clip. Finally, HMM model is utilized to calculate the likelihood of the topic vector, acting as an observation sequence of the model, so that classification of the music clip is reached. With the aim of making CTM model function effectively, density-based spatial clustering of applications with noise (DBSCAN) is used to cluster the words in the vocabulary. And according to the centers of the generated clusters, the number of topics K in CTM is determined. Because of the topic model