歌曲中相似片段的检测及其应用

发布时间：2018-02-09 16:11

本文关键词： 歌曲借用递归图状态空间嵌入相似片段检测相似度计算　出处：《复旦大学》2014年硕士论文　论文类型：学位论文

【摘要】：随着互联网的普及,几乎所有的音乐作品都对应的其数字化版本,且作品数量正飞速增长。追随着这种潮流,各大互联网服务供应商都推出了自己的在线歌曲库,内含海量歌曲作品。因此,怎样对大量的歌曲进行分类整理、信息提取,以便提供更好的检索、查询、欣赏等服务,成为近些年的热门研究课题。早期的歌曲库管理和歌曲检索功能依赖大量的人工标注,比如手工录入每首歌的年代、作者、流派等信息。这样的方法虽然可以保证信息标注的准确度,但成本非常大,且并不能满足很多常用的需求。由此,大量基于信号级别的歌曲信息处理相关的研究课题便大量涌现,如翻唱检索、哼唱检索、歌曲摘要提取、音频片段认证等。在所有这些研究课题中,歌曲相似片段检测与定位是一个非常重要的基础技术和子课题。音乐作品中相似片段是十分常见的：例如同一首歌曲中的副歌部分通常有着一样的旋律；又如同一首歌曲被多个歌手分别演唱所产生的不同版本。这里的相似可以是指某一特定方面的相似,如旋律、节奏、音色等。在很多情况下用人耳可以轻易辨别出相似的歌曲片段,然而对于计算机而言这个任务却很困难。由于直接基于歌曲信号提出的特征本身包含着大量的噪音,因此目前现有的相似片段检测算法有着诸多问题：首先,所有的算法准确度普遍不高；其次,每个算法都只能大致检测出两个给定片段的相似程度,但却不能精确定位；另外,现有算法无法检测出两段较短的相似片段。综上,目前的相似片段检测算法还有相当大的改进空间。为解决和改进上述提到的相关算法存在的问题,本文提出了一种新的相似片段检测算法。该算法吸收了现有的音乐信号处理研究中所用到的多种技术,如相似度矩阵、递归图、状态空间嵌入等,并在此基础上引入了数种创新技术,如局部阈值策略、基于递归图的相似片段提取技术等。本文提出的算法解决了先前类似算法中的多个缺陷：无法较准确定位相似片段位置；无法检测较短的相似片段。在此基础上,该算法还通过采用改进的阈值策略来提高相似片段检测的准确度。为检测该算法的有效性,本文将该算法应用于三个具体的系统中：改进的翻唱检索系统；基于该相似片段检测算法的歌曲摘要提取系统；歌曲借用检测系统。其中,歌曲借用检测是一个尚未有相关研究的课题。由于本文提出的算法可以检测歌曲中较短的相似片段,因此该课题得以第一次被进行尝试性的研究。针对上述三个系统的实验证明了本文所提出算法的有效性。
[Abstract]:With the popularity of the Internet, almost all music works correspond to their digital versions, and the number of works is growing rapidly. Following this trend, various Internet service providers have launched their own online song library. Therefore, how to sort out a large number of songs and extract information in order to provide better retrieval, inquiry, appreciation and other services, Early song library management and song retrieval functions depended on a large number of manual tagging, such as manual entry of each song, the author, This method can guarantee the accuracy of information tagging, but it is very costly and does not meet a lot of common needs. A large number of research topics related to song information processing based on signal level have emerged in large numbers, such as rap retrieval, humming retrieval, song abstract extraction, audio segment authentication and so on. Song similarity detection and location is a very important basic technology and sub-topic. Similar fragments are very common in musical works: for example, the chorus part of the same song usually has the same melody; It's like a different version of a song being sung by multiple singers. The similarity here can refer to the similarity of a particular aspect, such as melody, rhythm, Timbre and so on. In many cases, similar song fragments can be easily identified with the human ear, but this task is difficult for a computer. Because the features that are directly based on the song signal contain a lot of noise, Therefore, the existing similar segment detection algorithms have many problems: first, the accuracy of all algorithms is generally not high; secondly, each algorithm can only roughly detect the similarity of two given fragments, but can not accurately locate. The existing algorithms can not detect two short similar fragments. In summary, there is still considerable room for improvement of the current similar segment detection algorithms. In this paper, a new similar segment detection algorithm is proposed, which absorbs many techniques used in the research of music signal processing, such as similarity matrix, recursive graph, state space embedding, etc. On this basis, several innovative techniques are introduced, such as local threshold strategy, similar segment extraction based on recursive graph and so on. The algorithm proposed in this paper solves many defects in the previous similar algorithms: the location of similar segments cannot be located accurately; Short similar fragments can not be detected. On this basis, the improved threshold strategy is adopted to improve the accuracy of the detection of similar fragments. This paper applies the algorithm to three specific systems: an improved copy retrieval system; a song digest extraction system based on the similar segment detection algorithm; a song borrowing detection system. Song borrowing detection is a subject that has not been studied yet. Because the algorithm proposed in this paper can detect short similar fragments in songs, Therefore, this subject can be studied for the first time. Experiments on the above three systems have proved the effectiveness of the proposed algorithm.
【学位授予单位】：复旦大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.2

【相似文献】