感谢华东师大艺术系硕士研究生徐进提供并整理该文资料---洪啸音乐教育工作站

学位授予单位 清华大学  
中文关键词 音乐识别;特征提取;STFT;调式自动判别;匹配比较  
导师 任勇  
论文级别 硕士  
学科专业名称 信息与通信工程  

 
中文摘要 音乐的计算机自动识别是新兴的交叉学科,其研究涉及物理学、信号处理、人机交互、音乐理论和音乐心理学等诸多学科知识。其主要任务是通过对音频信号的处理和特征提取,获取音乐内容的相关信息,进而用于比较、分类乃至自动录谱等。本文研究的钢琴音乐演奏的计算机识别正是将计算机多媒体技术、信号处理与模式识别的相关知识和技术同音乐理论相结合,用计算机模拟人对音乐认知分析过程,分析解析音乐演奏并评价演奏的优劣。 本文工作属于音乐识别的理论研究及其应用范畴,通过机器对音频信号的自动分析及处理,实现了对钢琴音乐演奏的特征识别与正误判定。钢琴演奏音乐的机器识别,可从一定程度上辅助钢琴演奏的教学。 论文首先提出钢琴演奏音乐正误识别系统的概念与结构,主体工作围绕音乐识别相关理论、技术实现与实验测试展开。根据音乐理论和音的物理属性,提出可用于识别的音乐特征,扩展以往的识别特征集。在完成对表征音乐特性的七项音乐特征的提取研究和实现的同时,深入音高时值、调式调性特征提取单元的研究。对比时域并行处理、谐波峰值法和小波变换后,改进基频提取方法,并首先给出时间-事件表达方式,用来表示提取的音高时值信息。实验证明,这种基于短时傅里叶分析(STFT)的基频提取方法有良好的识别率。论文工作首次实现了调式调性的计算机自主识别,完成机器对调式调性风格的分析,使独立的单音性质与乐曲整体感觉结合起来。在特征提取基础上,本工作实现了基于机器的音频特征与乐谱特征的比较,以及计算机对钢琴演奏音乐的打分、评定与图像显示。 钢琴演奏音乐正误的计算机识别系统很好地综合了音乐和技术领域的知识,初步实现了机器自动判别演奏准确性的功能,使自动识别在辅助音乐教育等领域得到良好发挥。经实验测试,系统的整体性能达到预期希望并满足一定的实用要求,为今后这方面的研究奠定基础。  

 
英文摘要 Music recognition is a rising interdisciplinary research field, which involvesphysics, signal processing, human-computer interaction, music theory and musicpsychics. The goal of this recognition is to extract the music characters by audiosignal processing and feature selection so that they can be used in music analysis,classification and automatic score recording. To describe piano music characters bycomputer, this dissertation combines the technique of multimedia, signal processingand pattern recognition with music theory to make the computer to imitate the pianoperformance evaluation of human. The work in this paper belongs to the research field of music identificationtheory and its application. Employing audio signal analysis and processing, we makeit possible for computer to recognize music characters and evaluate its performance.We first outline the conception and framework of automatic piano performanceevaluation system based on music recognition theory. The evaluation system is thenimplemented with performance testing in this paper. Based on music theory and thenotes physical attributes, we propose seven music features for identification. Amongthese seven features, we focus on pitch, duration and tonality characters detection.Experiments are conducted for comparisons among time-domain parallel processing,harmonics peak value method and wavelet transform. The paper advances the pitchdetection algorithm and for the first time illustrates the time-event figure to explainthe result of pitch and duration features recognition. Experimental results show thatthe advanced STFT pitch detection method achieves high recognition ratio. The paperfor the first time implements the computer detection of the music piece tonality,integrating the property of each note with the style of the whole piece. Based onmusic feature selection, the system is able to evaluate the piano playing performanceby scores and visualization. The automatic piano performance evaluation system wellsynthesizes the knowledge in both music and technology areas, making it possible forcomputers to evaluate music performance automatically. Intelligent music recognition will help the music performance learning. The experiment shows that the systemfulfills the anticipated requirements, therefore establishes solid foundation for thefuture work.  
DOI CNKI::CDMD:10003.2.2005.6977