语境化音乐流派识别

发布时间：2018-01-29 17:21

本文关键词： 音乐流派分类声学特征语境化标签特征堆叠泛化　出处：《安徽大学学报(自然科学版)》2017年03期 　论文类型：期刊论文

【摘要】：音乐流派识别(music genre recognition,简称MGR)是一项应用前景广阔的研究领域,同时也能够为音乐机器学习技术提供一个良好的测试台.在过去几年中,涌现出了大量基于机器学习的MGR系统.然而,单纯依靠底层声学特征的MGR算法在分类时往往不能得到令人满意的结果.作者受到语境化机器学习技术的启发,提出了基于堆叠泛化的MGR分类系统.该系统通过提取互联网中用户标注音乐时使用的标签,同时融合声学特征完成语境化音乐流派识别.在论文中,MGR系统分为内容层和语境层,内容层由提取Mel倒谱系数(Mel frequency cepstrum coefficient,简称MFCC)和快速傅里叶变换(fast Fourier transformation,简称FFT)特征的声学特征表示,语境层由额外标签表示.在数据集GTZAN的实验结果表明,在数据集GTZAN下,语境化的MGR分类准确率远高于基于声学特征的MGR分类准确率,实现了20%的分类性能的提升.
[Abstract]:Music genre recognition (music genre recognition, referred to as MGR) is a research field wide application prospect, but also can provide a good technical test bench for music machine learning. In the past few years, the emergence of a large number of machine learning based MGR system. However, the MGR algorithm only depend on the underlying acoustic features when the classification can not get satisfactory results. The author has inspired context of machine learning technology, put forward the MGR classification system based on stacked generalization. The system user through Internet Music Label extraction using the label, and acoustic features complete context fusion music genre recognition. In this paper, the MGR system is divided into the content layer and content layer by layer context, Mel cepstrum extraction (Mel frequency cepstrum coefficient, referred to as MFCC) and fast Fourier transform (fast Fourier transformation, Jane The FFT is the representation of the acoustic features of the feature, and the context layer is represented by the extra labels. The experimental results in the dataset GTZAN show that under the dataset of GTZAN, the accuracy of the contextual MGR classification is much higher than that of the MGR classification accuracy based on the acoustic features, achieving a 20% classification performance improvement.

【作者单位】：曼彻斯特大学计算机科学学院;
【分类号】：J611.1;TP181
【正文快照】： 随着多媒体技术的持续发展,数字音乐在包括无线电广播和因特网的各种多媒体中的应用越来越广泛.与此同时,随着数字音乐在因特网的快速普及,对于音乐处理与操作服务的需求也在上升.在这种条件下,针对如何有效搜索与管理数字音乐的工具的研究就受到了越来越多的关注.音乐流派识

【相似文献】