【作者】 唐磊；

【导师】 李学庆；

【作者基本信息】 山东大学 ， 计算机软件与理论， 2012， 博士

【摘要】 音乐是人们生活中不可缺少的一部分。随着互联网技术的迅速发展和压缩存储技术的成熟,人们收集音乐、存储音乐的能力得越来越强,但对音乐的分析、处理能力却没有得到相应的提升。目前对音乐的分析和管理主要有以下几种方法：一是基于标签的方法,根据艺术家、专辑、音乐流派等附加信息对音乐进行分析、过滤；二是基于内容的方法,它以音乐本身的采样信息为基础,从中提取节奏、音色、音高等多种音乐特征,通过对音乐特征的相似性分析,找出隐藏的规律和现象,帮助人们更好的理解音乐。三是基于机器学习的方法,通过对音乐样本的学习掌握对应关系,实现音乐的分类。然而,这几种方法大都以音乐内容为基础,缺少对用户行为和偏好的分析,无法满足不同用户的欣赏需求。虽然部分学者提出了基于情感和内容的音乐分析方法,将主观因素和客观信息相结合,但情感仅仅是对人们主观感受的一种概括和总结,种类有限,还是无法体现个体用户之间的差别。另外,对音乐的描述和音乐关系的展现也是音乐分析过程中的一个障碍,虽然有学者提出了基于可视化的解决方案,但还存在一些局限性和不足,没有对音乐和音乐关系进行深入的挖掘,层次划分粒度不够。本文充分调研了音乐分析的国内外研究现状,针对音乐分析过程中涉及的若干关键问题和研究难点,以音乐内容和用户交互历史为研究对象,以可视化和分析为主要方法,提出了基于内容和用户历史的音乐可视分析。本文的主要研究内容和创新点如下：1.音乐特征的提取和优化。本文充分研究了音乐信息提取方面的相关文献,以音乐内容为基础,提取了Timbre, Rhythm和Chroma三种特征作为主要分析依据。三种特征组成的多维向量较为复杂,且存在冗余信息,对分析效率影响较大。针对这一问题,本文提出了基于可视化技术的特征优化方法,利用扩展的平行坐标轴消除作用较小的特征分量,利用基于维密度和聚类的散点图消除作用相似的冗余特征。实验结果证明,本文方法能够有效解决分析精度和分析效率问题。2.音乐推荐是音乐分析的一个研究难点和热点,当前的音乐推荐算法由于缺少对主观因素的分析,无法体现个体用户的欣赏偏好。针对这一问题,本文提出了基于内容和用户历史的音乐推荐算法,利用协作推荐算法分析用户行为,利用基于图的分析方法和相似性分析方法分析音乐特征,最后将多种算法有效融合。实验结果表明,本文设计的推荐算法相比传统算法准确度更高,能够针对不同用户推荐其可能喜欢的音乐,有效解决了用户欣赏偏好和个体差异问题。3.对音乐及音乐关系的描述和展现是音乐分析的又一难点,当前的研究成果通常使用Ove rView+Detail技术或者Focus+Contex技术对全局信息和局部细节进行描述,但在音乐重要性和关联性分析方面的研究还不够充分。针对这一问题,本文提出了基于层次的音乐信息可视化,根据用户关注度和音乐间的相似关系将音乐划分为重要音乐、次重要音乐和辅助音乐三个层次。该方法利用推荐算法和布局技术实现了重要音乐和次重要音乐的展现。该方法还提出了音乐云的概念,利用分段高斯方程和可视编码技术,以云片的形式对辅助音乐和相应关系进行了可视化描述。用户调查结果显示,大部分用户对基于层次的音乐信息可视化方法较为满意,认为层次清晰,符合认知过程。特别是音乐云的设计给用户带来了全新的感受,大部分用户表示音乐云的设计对决策制定很有帮助。4.在上述研究内容和工作的基础上,本文设计并实现了一个音乐可视分析的原型系统,并对音乐信息的可视编码、播放列表的生成和一些常用的交互技术能进行了详细介绍。更多还原

【Abstract】 Music plays a key role in our daily life. With the rapid development of internet and the great improvement of store technology, now people have got incredible ability to collect and store music. However, their ability to deal with music does not scale with the music library. To address this problem, several approaches have been proposed to analyze music. One method is tag-based, which analyzes and filters music by artist, album, genre and other tag information. The second approach is content-based. According to music content, several features can be extracted and used to analyze the similarity between music, such as timbre, chroma and rhythm. With the content-based method, people can easily find out the rules, relations and patterns hidden in music features. Machine learning is another music analysis approach. Through learning the unknowns in samples, all the music can be divided into several groups. In a word, all these approaches are based on music content without taking account users’preference and behavior. To deal with this problem, several researchers proposed the content and mood based solution. They use both content and people’s mood to analyze music, and try to make everyone satisfied. However, only several kinds of mood are not enough. In addition, how to describe music and how to illustrative the relative relationship is another unsolved problem. Though some researchers proposed visualization-based solution, their methods still have some limitations, and can’t reveal more details.This thesis first surveys the existing work on music analysis. Based on this survey, we focus on several key problems and propose content and user history based music visual analysis. The innovation and contribution of this thesis mainly include:1. Based on music information retrieval, this thesis first extracts three music features-Timbre, Chroma and Rhythm as the basis of our work. These features include some redundant and unnecessary information, which affects the analysis efficiency. To deal with this problem, two visualization-based music feature optimization approaches are proposed. An improved parallel coordinates is used to eliminate the unnecessary information, while the extending scatterplots technique is used to explore the redundant feature. The experiment shows that these approaches are feasible and work well. 2. Music recommendation algorithm is a key point in music analysis. However the existing work only focuses on music content without taking account users’preference too much. As a result, these algorithms fail to reveal users’ preference. To address this problem, a content and user history based music recommendation algorithm is proposed. The collaborative algorithm is used to analyze a user’s interaction history, while the similarity algorithm and graph-based algorithm are used to analyze the music features. The experiment indicates that our approach performs better than the traditional algorithm, and recommends suitable music list to different user according to his preference.3. The description of music and music relationship is another focus in music analysis. According to the existing approaches, Overview+Detail and Focus+Contex techniques are usually used to describe the overall information and details. However, these approaches have the disadvantages of music importance and relationship illustration. This thesis proposes a layer-based music information visualization to solve this issue. In this approach, all pieces of music are divided into three layers-most important layer, secondary important layer and auxiliary layer according to the user’s preference and similarity between music. With recommendation algorithm and layout technique, we complete the visualization of the first two layers. While with music cloud, we reveal the auxiliary music and corresponding relationship. The user study shows that most of the users are satisfied with the layer-based music information visualization, and think this approach is intuitive. In addition, most of the users are interested in the design of music cloud, and feel it’s helpful to make a decision.4. Finally, a prototype system for music visual analysis is designed and implemented based on the proposed algorithms and our work. Especially, several relative techniques are introduced separately as a footnote, such as visual coding, playlist creation and some interaction techniques.更多还原