单声道的歌声分离方法研究

发布时间：2018-12-18 18:16

【摘要】：单声道歌声分离是基于内容的音乐分析领域中的重点和难点。本文从伴奏和歌声两个角度出发,总结了当前主流的歌声分离算法并提出改进。本文的主要研究内容和创新点如下:1、本文从伴奏和歌声两个角度,全面总结了当前主流的单声道歌声分离方法,并将其划分为基于伴奏重复模式、基于歌声基频估计两类。在综述上述两类方法时,本文还分析指出各类方法中的主要思想和存在的约束限制。通过综述发现,伴奏重复模式和歌声谐波结构是当前歌声分离方法中最重要的两个线索,本文后续将从这两个方面选出参考方法加以研究改进。2、本文在详细分析基于伴奏重复模式的歌声分离方法后,选择约束条件少、计算量小的最近邻中值滤波(Nearest Neighbors Median Filtering,NNMF)歌声分离作为参考方法加以深入研究。本文分析了 NNMF方法存在的不足,即不适合处理歌声能量强的乐曲,并提出多子带划分结合NNMF的歌声分离方法。当混合音乐信号歌声能量较强时,通过多子带的划分能使一些子带中伴奏能量较强从而提升歌声分离性能。本文在当前主流的测试样本集MIR-1K上全面客观评测了参考方法和改进方法。结果显示,改进方法分离歌声和伴奏的整体评价指标都比参考方法好,三种信噪比的测试样本集下,分离歌声整体评价指标至少提高2.5dB,分离伴奏整体评价指标至少提高0.8dB。3、本文在详细分析基于歌声基频估计的歌声分离方法后,选择理论框架完善、无谐波重叠问题的瞬时混合模型(Instantaneous Mixture Model,IMM)歌声分离方法作为参考方法加以深入研究。本文分析了 IMM方法存在歌声基频估计过于简单粗略的问题,并引入歌声检测和基于旋律先验统计的Melodia歌声基频估计算法加以改进。本文在测试样本集MIR-1K上全面客观评测了参考方法和改进方法,结果显示,引入更加合理的歌声基频估计算法Melodia之后,分离性能得到较大提升,分离歌声整体评价指标至少提高1.67dB,分离伴奏整体评价指标至少提高近1dB。引入歌声检测之后,分离歌声中包含的伴奏分量更少,但由于歌声检测性能的限制,歌声分离性能只在-5dB测试样本集下有所提升。测试样本集的客观评测结果说明本文对参考方法的讨论分析和探索改进是合理有效的。
[Abstract]:Mono-channel song separation is an important and difficult point in the field of content-based music analysis. In this paper, from the two angles of accompaniment and singing, we summarize the current mainstream singing separation algorithm and propose some improvements. The main contents and innovations of this paper are as follows: 1. From the two angles of accompaniment and singing, this paper summarizes the current mainstream methods of mono song separation, and divides them into two types: one is based on accompaniment repetition mode, the other is based on the estimation of the fundamental frequency of singing. In this paper, the main ideas and constraints of these two methods are also analyzed. It is found that the repetition pattern and the harmonic structure of the accompaniment are the two most important cues in the current singing separation methods. This paper will select the reference methods from these two aspects to study and improve. 2. After detailed analysis of the song separation method based on accompaniment repetition pattern, this paper chooses the nearest median filter (Nearest Neighbors Median Filtering,NNMF) as the reference method with less constraint and less computation. In this paper, the shortcomings of NNMF method are analyzed, that is, it is not suitable to deal with music with strong singing energy, and a method of song separation based on multi-subband division and NNMF is proposed. When the singing energy of the mixed music signal is strong, the performance of song separation can be improved by the division of the multi-subbands to make the accompaniment energy stronger in some sub-bands. This paper gives a comprehensive and objective evaluation of the reference method and the improved method on the current mainstream test sample set MIR-1K. The results show that the whole evaluation index of the improved method for separating singing and accompaniment is better than that of the reference method. Under the three SNR test samples, the overall evaluation index of the song separation is improved by at least 2.5 dB. The overall evaluation index of the separation accompaniment is at least 0.8dB.3. after the detailed analysis of the singing separation method based on the fundamental frequency estimation of the song, this paper chooses the instantaneous mixed model (Instantaneous Mixture Model, which has a perfect theoretical framework and no harmonic overlapping problem. IMM) the method of song separation is studied as a reference method. In this paper, we analyze the problem that the IMM method is too simple and rough to estimate the fundamental frequency of singing, and introduce the song detection and the Melodia estimation algorithm based on the prior melodic statistics to improve it. In this paper, the reference method and the improved method are evaluated objectively on the test sample set MIR-1K. The results show that the separation performance is greatly improved by introducing a more reasonable fundamental frequency estimation algorithm (Melodia). At least 1.67 dB was increased in the whole evaluation index of the separated singing and at least 1 dB was increased in the whole evaluation index of the separation accompaniment. After the introduction of song detection, there are fewer accompaniment components in the separated songs, but due to the limitation of the performance of song detection, the performance of song separation is only improved under the-5dB test sample set. The objective evaluation results of the test samples show that the discussion, analysis and improvement of the reference method in this paper are reasonable and effective.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.3

【相似文献】