音乐情感参数化系统的研究与实现

发布时间：2018-08-26 21:11

【摘要】：在当今互联网浪潮的推动下，数字音乐的数量出现了爆炸式的增长，急需高效的分类管理方法。近年来，国内外学者针对音乐检索展开了广泛、深入的研究，但是未能取得广泛的应用，一方面，，音乐检索是一个多学科交叉领域，研究难度大；另一方面，目前的众多研究多以音乐流派和情感标签作为分类目标，类似传统的分类管理方式，存在局限性。因此，开展音乐检索相关研究具有重要的研究价值。针对目前基于情感的音乐检索研究的不足，本文提出以参数来表示音乐情感强弱的方法，首先提取音乐情感特征，组成特征向量，然后利用fisher算法进行维数压缩，再通过大量的音乐样本对音乐情感参数化系统进行训练，最终得到节奏、音调和音色三个描述音乐情感强弱的参数。本文的研究成果主要有以下几个方面：首先，音乐情感特征的研究，通过实验证明MFCC是一组非常重要的参数，它在很大程度上决定了音乐情感分类的正确率。对于MFCC特征维数的选取，实验结果表明，13、14维是比较合理的。不同特征之前没有相互排斥，而是相互补充，因此搭配使用不同的特征有助于提高总体的分类正确率。其次，Fisher和SVM两种不同算法分类性能比较，在音乐情感类别很少的情况下，比如2个类别，两者分类性能接近，为了方便分类器设计、节省计算资源，优先选择Fisher分类器；在类别很多的时候，为了保证分类正确率，应该选择SVM这一类基于机器学习理论分类器；当类别特别多，起到关键作用的是音乐情感特征的选取，而不是分类器算法，应该将研究重点放在这方面。最后，音乐情感参数化系统的设计，本文以Marsyas音频处理库为基础，搭建了基于数据流模型的系统框架，选择了适当的情感特征组成特征向量，同时选择Fisher算法作为分类器，使用大量的音乐样本进行了系统训练，并对节奏、音调和音色三个参数进行参数归一化处理，最终完成了音乐情感参数化系统的实现。测试实验结果表明，本文实现的系统能够达到88%的识别正确率，基本满足实际应用需求，可以为相关的音乐管理软件提供搜索引擎，促进音乐自动搜索技术的发展。
[Abstract]:The number of digital music has been increasing explosively under the impetus of the current Internet wave, and the efficient classification management method is urgently needed. In recent years, scholars at home and abroad have carried out extensive and in-depth research on music retrieval, but they have not been widely used. On the one hand, music retrieval is a multidisciplinary field, which is difficult to study; on the other hand, At present, many researches take music genre and emotion label as the classification target, similar to the traditional classification management, there are some limitations. Therefore, the development of music retrieval related research has an important research value. Aiming at the deficiency of the research on music retrieval based on emotion at present, this paper proposes a method to express the intensity of music emotion by parameter. Firstly, the feature vector of music emotion is extracted, then the dimension is compressed by fisher algorithm. Then through a large number of music samples to carry on the training to the music emotion parameterization system, finally obtains the rhythm, the tone and the timbre to describe the music emotion strong and weak parameter. The main research results of this paper are as follows: firstly, the research of music emotion characteristics proves that MFCC is a group of very important parameters, which determines the correct rate of music emotion classification to a great extent. For the selection of MFCC characteristic dimension, the experimental results show that 1314 dimension is reasonable. Different features are not mutually exclusive but complement each other before, so collocation of different features can improve the overall classification accuracy. Secondly, the classification performance of SVM and SVM are compared. In the case of few categories of music emotion, such as two categories, the classification performance is similar. In order to facilitate the design of classifier and save computing resources, Fisher classifier is chosen first. When there are many categories, in order to ensure the classification accuracy, we should choose SVM, which is based on machine learning theory classifier, and when there are many categories, it is the selection of music emotion feature, not the classifier algorithm that plays a key role. Research should be focused on this area. Finally, the design of the music emotion parameterization system, based on the Marsyas audio processing database, the system frame based on the data flow model is built, and the appropriate emotion features are selected as the feature vector, and the Fisher algorithm is chosen as the classifier. A large number of music samples are used for systematic training, and the parameters of rhythm, tone and tone color are normalized. Finally, the realization of music emotion parameterization system is completed. The test results show that the system can achieve 88% correct recognition rate, basically meet the practical needs, can provide a search engine for the related music management software, and promote the development of automatic music search technology.
【学位授予单位】：华南理工大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TN912.3

【参考文献】