音乐和弦识别的研究

发布时间：2018-03-13 09:18

本文选题：和弦识别　切入点：对数音级轮廓特征　出处：《天津大学》2016年博士论文　论文类型：学位论文

【摘要】：随着互联网带宽的增长,以及多媒体信息压缩技术的不断发展,互联网上数字音乐的存储和发布越来越普遍。为了应对用户随时随地检索的需求,基于内容的音乐检索应运而生。MIR中的中层特征就包括和弦,它包含了大量能够表现音乐属性的信息,对于分析音乐结构和旋律方面具有非常重要的作用。因此,本文针对音乐和弦识别进行了深入的研究,提出了鲁棒性音乐和弦识别特征和两种和弦估计方法。本文综合应用部分乐理、信号处理、模式识别等相关知识,提出了序列化稀疏表示分类和序列化支持向量机的和弦识别方法。其主要研究内容是以信号处理为基础,从特征提取和和弦估计两方面研究和弦识别。主要完成的工作包括以下几个方面:(1)提出了鲁棒性对数音级轮廓特征。和弦识别的一个关键是特征,在基于节拍的基础上提出了LPCP,使得LPCP能够更好地表达音频内容,提高和弦识别率;同时为了尽可能降低歌声的影响,在计算PCP前,对音频文件进行歌声伴奏分离,使得伴奏能够更好地包含和弦特征,这样音频文件对和弦识别具有更好的鲁棒性;(2)本文提出了基于序列化稀疏表示分类器的音乐和弦识别方法。在稀疏表示分类中,建立和弦样本数据库,对输入的音频片段进行和弦估计。在此基础上,结合隐形马尔科夫链模型,克服需要大量训练得到模型参数的缺点,提出序列化稀疏表示模型。在对MIREX’09的数据库中的大小和弦识别时,本论文提出的方法在使用本文的特征进行识别时,识别率均高于目前的识别方法。(3)提出了序列化支持向量机的音乐和弦识别方法。为了克服稀疏表示分类时间较长的缺点,引入支持向量机用于和弦识别。该模型只需要提前训练好参数,用于和弦估计时间较短。同时结合音乐和弦在时域上的变化特点,进一步改进支持向量机,提出序列化支持向量机模型。
[Abstract]:With the growth of Internet bandwidth and the continuous development of multimedia information compression technology, the storage and distribution of digital music on the Internet is becoming more and more common. Content-Based Music Retrieval (CBIR) emerges as the times require. The middle level features of music retrieval include chords, which contain a large amount of information that can express the musical properties, and play a very important role in analyzing the music structure and melody. In this paper, the characteristics of robust music chord recognition and two kinds of chord estimation methods are proposed, and some related knowledge, such as music theory, signal processing, pattern recognition and so on, are synthetically applied in this paper. In this paper, a method of serialized sparse representation classification and serialization support vector machine is proposed, which is based on signal processing. This paper studies the recognition of chords from two aspects: feature extraction and chord estimation. The main work accomplished includes the following aspects: 1) the robust logarithmic tone level contour feature is proposed. One of the key points of chord recognition is the feature. In order to reduce the influence of singing as much as possible, LPCP is used to separate audio files before calculating PCP, which makes LPCP express audio content better and improve the recognition rate of chords. In this paper, a method of music chord recognition based on serialized sparse representation classifier is proposed. The chord sample database is established, and the input audio segment is estimated by the chord. On this basis, combined with the invisible Markov chain model, it overcomes the shortcoming that a lot of training is needed to obtain the model parameters. A serialized sparse representation model is proposed. When recognizing the size and chord in the MIREX'09 database, the method proposed in this paper uses the features of this paper to recognize. The recognition rate is higher than that of the current recognition method. (3) Serialization support vector machine (SVM) is proposed to recognize music and chord. In order to overcome the disadvantage of long time of sparse representation classification, the method of serialized support vector machine (SVM) is proposed. Support vector machine (SVM) is introduced for chord recognition. The model only needs to train the parameters in advance, and the estimation time of chord is short. At the same time, the support vector machine (SVM) is further improved according to the changing characteristics of music chord in time domain. A serialization support vector machine model is proposed.
【学位授予单位】：天津大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TN912.3

【相似文献】