基于最大最小距离法的音乐节拍跟踪算法研究

发布时间：2018-09-06 15:22

【摘要】：随着因特网与网络技术的快速发展,人们可以接触到大量的在线音乐数据,比如音乐原声信号、歌词、音乐曲风或者内容的分类以及其他网络用户的歌单等等。这种科技的进步让用户在听音乐时有了越来越多的乐趣,同时,也对数据的处理提出了更高的要求,如何让计算机更好地丰富用户的音乐体验成为一个热门问题,也促进了音乐信息检索领域的深入研究。音乐信息检索,是一个跨学科的研究领域,涉及到音乐学、心理学、音乐学术研究、信号处理、机器学习等等。节拍跟踪是音乐信息检索的基础问题之一。人们会不自主的跟随音乐跺脚或者点头的过程称为节拍跟踪,计算机的节拍跟踪算法正是对人类这一感知过程的模拟。过去的二十多年中,节拍跟踪研究领域已有大量深入的研究,也有越来越多的节拍跟踪算法应用于实际生活中。本文在认真研究节拍跟踪相关研究成果的基础上,结合音乐基本理论与音频信号技术,提出一种基于最大最小距离法的节拍跟踪算法,核心为起始节拍点的确定、BPM特征值提取和有效峰值提取三部分。本文的创新点在于将聚类算法应用于节拍跟踪研究,将峰值提取问题抽象为分类问题,从聚类的角度完成节拍序列的提取。具体研究步骤可概括为以下几个方面:首先,对音乐信号进行预处理,统一采样频率及幅度范围。提取音乐信号的1-2s片段进行时域处理,通过分析该片段的能量谱变化,确定起始节拍点。其次,对音乐信号进行短时傅里叶变换得到频谱,根据人类听觉系统的感知特性,对频谱幅度进行对数处理,通过半波整流输出端点强度曲线及其峰值的相位信息。根据端点强度曲线的自相关特性提取BPM特征值。最后,根据音乐的节拍和速度的关系以及周期信号的性质,利用最大最小距离法对端点强度曲线的峰值点进行有效聚类,输出节拍序列。本文采用MIREX2006测试数据进行实验,将本文的提出算法与MIREX2013节拍跟踪比赛中性能较好的算法进行对比。实验结果表明,本文提出的基于最大最小距离法节拍跟踪算法对于不同曲风、不同节奏类型的音乐信号,四项算法评估指标P-Score、Cemgil、CMLc和AMLt的均值分别为57.35510、38.70537、17.15240和47.25912,能准确有效地检测出节拍序列,在全局正确性、连续正确率两方面都有较大优势,综合性能稳定。
[Abstract]:With the rapid development of Internet and network technology, people can get access to a large number of online music data, such as music soundtrack, lyrics, music style or content classification, as well as other network users' song list and so on. The advances in technology have made it more and more fun for users to listen to music. At the same time, they have put forward higher requirements for the processing of data. How to make computers better enrich the music experience of users has become a hot issue. It also promotes the in-depth research in the field of music information retrieval. Music Information Retrieval is an interdisciplinary research field involving musicology, psychology, music academic research, signal processing, machine learning and so on. Beat tracking is one of the basic problems in music information retrieval. The process of stamping or nodding with music involuntarily is called rhythm tracking, and the computer beat tracking algorithm is the simulation of this process of human perception. In the past twenty years, there have been a lot of in-depth researches in the field of beat tracking, and more beat tracking algorithms have been applied in real life. In this paper, a beat tracking algorithm based on the maximum and minimum distance method is proposed, based on the research results of rhythm tracking, combined with the basic theory of music and audio signal technology. The core is the determination of the starting beat point and the extraction of the BPM eigenvalue and the effective peak value. The innovation of this paper lies in applying the clustering algorithm to the research of beat tracking, abstracting the peak extraction problem as a classification problem, and completing the extraction of beat sequences from the point of view of clustering. The specific research steps can be summarized as follows: firstly, preprocessing the music signal and unifying the sampling frequency and amplitude range. The 1-2s segment of music signal was extracted and processed in time domain, and the starting beat point was determined by analyzing the energy spectrum change of the segment. Secondly, the spectrum of music signal is obtained by short-time Fourier transform. According to the perception characteristics of human auditory system, the amplitude of spectrum is processed logarithmically, and the intensity curve of endpoints and the phase information of peak value are output by half-wave rectifier. The BPM eigenvalues are extracted according to the autocorrelation characteristics of the endpoint strength curve. Finally, according to the relationship between the rhythm and the speed of music and the property of the periodic signal, the maximum and minimum distance method is used to cluster the peak points of the endpoint intensity curve effectively, and the beat sequence is outputted. In this paper, the MIREX2006 test data are used to carry out the experiment, and the proposed algorithm is compared with the algorithm with better performance in the MIREX2013 beat tracking competition. The experimental results show that the proposed beat tracking algorithm based on the maximum and minimum distance method is suitable for different music signals with different styles and different rhythms. The average values of P-Scorex Cemgilg CMLc and AMLt are 57.3551010 ~ 38.70537 ~ 17.15240 and 47.25912 respectively, which can accurately and effectively detect the rhythm sequence, and have great advantages in both global correctness and continuous accuracy, and the comprehensive performance is stable.
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN911.7;TP18

【参考文献】