【作者】 孙健；

【导师】 孙怡；

【作者基本信息】 大连理工大学 ， 电子与通信工程， 2011， 硕士

【摘要】 音乐是人类最伟大的艺术之一,它深远地影响人类的精神生活。近年来,由于数字音频处理技术及多媒体技术快速发展,数字音乐已经深入到人们生活的方方面面,不断影响着人类欣赏音乐的方式,人们对音乐欣赏的需求不断增长。与此同时,数字音乐信息获取技术得以快速发展,国内外很多科研工作者和研究机构对数字音乐特征获取和音乐内容理解进行了深入的研究。为促使音乐信息获取技术的发展及其相关技术的应用,本文对基于数字音频信号的音乐节奏信息获取技术进行了广泛的调研与深入的研究,设计了一种实时解析音乐节拍点,跟踪音乐速度与拍子结构等节奏信息变化的系统,目的在于使计算机也能够和人样“听懂”音乐的韵律节奏。在设计节奏信息获取系统的过程中,本文首先对原始音频信号进行特殊的简化处理,以得到音乐事件起始点检测函数,继而以此函数序列作为节奏获取模型的输入,利用基于无偏的自相关函数与梳状抽样函数模板的方法估计音乐节拍点周期与节拍点相位,并跟踪音乐速度的变化与识别拍子结构信息。但这种实时预测模型对音乐速度的变化比较敏感,预测的节拍点位置往往不够准确,导致模型可靠性降低。为提高音乐节奏实时预测模型的可靠性与准确性,本文利用基于隐马尔科夫模型的Viterbi解码算法与动态规划的方法对历史音频数据进行分析,求解出最佳的音乐节拍点位置序列,并以此对实时预测模型进行改进。最后本文利用已进行准确值标注的音乐音频片段对系统进行了实验与评估。实验和评估的结果表明改进后的模型达到了提高预测准确性的预期目的。更多还原

【Abstract】 Music is the one of the greatest arts. It has a profound impact on the activity of human spirit. In recent years, digital music has become deeply diffused among the diverse aspects of human activity owing to the rapid developments of digital audio processing technology and multimedia technology, which continuously affects the manner of the music appreciation while it increases the demand for the music appreciation. In the meantime, digital music information retrieval (MIR) technology has been well developed. At home and abroad, more and more scientific researchers and institutes carried out research works in depth on musical characteristics retrieval and musical content understanding.To promote the growth and development of MIR as well as its relative technology applications, we executed an extensive investigation and made a deep research on it. We have already designed a real-time system capable of predicating beats and tracking musical rhythmic information such as tempo and meter with the purpose of enabling the computer to hear the rhythmic structure of music as human beings do.In the process of designing our system, we firstly preprocessed specially the raw audio signal so as to obtain the onset detection function of musical events, and then took this function sequence as the input into the musical rhythmic retrieval model. Utilizing the method based on the unbiased autocorrelation function and comb sampling function template, we estimated the beat period and beat phase, tracking tempo and recognizing meter at the same time. Because of the sensitivity to big tempo change, the accuracy of the predicted beats was often unsatisfied, leading to a low reliability of the system. In order to increase the accuracy and reliability of real-time prediction model, we devised an approach of analyzing the past audio data using Viterbi decoding algorithm and a method based on dynamic programming to deduce the optimal beat sequence. After that, we improved the real-time prediction model of musical rhythmic information.Eventually, we carried out experiments and evaluated our system by several musical audio experts labeled by ground truth. The results turn out that the improved model has achieved the anticipated goal of increasing prediction accuracy.更多还原