面向巴松演奏音乐的精准音频乐谱比对方法研究  

Research on accurate audio-to-score alignment method for bassoon music

作　　者：连志成 程皓楠 张加万[1] LIAN Zhicheng;CHENG Haonan;ZHANG Jiawan(College of Intelligence and Computing,Tianjin University,Tianjin 300350,China;State Key Laboratory of Media Convergence and Communication,Communication University of China,Beijing 100024,China)

机构地区：[1]天津大学智能与计算学部,天津300350 [2]中国传媒大学媒体融合与传播国家重点实验室,北京100024

出　　处：《中国传媒大学学报（自然科学版）》2022年第2期54-62,共9页Journal of Communication University of China：Science and Technology

基　　金：国家自然科学基金项目(62172295);国家重点研发计划项目(2019YFC1521200)。

摘　　要：音频乐谱比对技术是一种将音频音乐与对应乐谱符号进行对齐的技术,是音乐信息检索(MIR)领域的重要研究方向。由于巴松的器乐、演奏、曲式特点,现有的音频乐谱比对方法无法精准处理巴松音频乐谱对齐任务。本文提出了一种面向巴松演奏的由粗到精、逐层细化的分段式高精度音频乐谱比对方法,针对巴松演奏音乐构建了首个由巴松独奏音频和对应乐谱组成的多曲式分类BSAMS(Bassoon-Solo-Audio-Midi-Score)数据集,并手工标注了音符起始时间和音符对应关系。具体来说,首先基于动态时间规整和音符起始点检测,设计了一种基准点和候选点生成算法,实现了音频乐谱对齐的粗略估计。其次,提出了一种基于支持向量机模型的音频乐谱点对筛选算法。最后,为了对精准匹配的对齐结果进行校验修正,提出了一种基于音乐理论的匹配修正算法,从而进一步提升了比对的准确度。在BSAMS数据集上对不同类型音乐进行实验,结果表明,本文提出的方法相比于传统通用音频乐谱比对方法可以达到在准确度上平均提高32.5%。Audio-to-score alignment is a technology that aligns music audio with its corresponding score symbols,which is important in the field of music information retrieval(MIR)and has important practical significance for the development of music analysis and other fields.Due to the characteristics of bassoon′s instrument,performance,and musical forms,the existing audio-to-score alignment methods cannot accurately complete the task of bassoon′s audio-to-score alignment.To solve the above problems,we propose an accurate audio-to-score alignment method for bassoon music.First of all,we take the lead in constructing the BSAMS(Bassoon Solo Audio Midi Score)dataset composed of musical forms classified bassoon solo audio and corresponding scores for bassoon music,and manually annotate the onset of the notes and the correspondence of the audio and score.In order to achieve high-precision audio-to-score alignment,based on the BSAMS dataset,we design a segmented accurate audio-to-score alignment method from coarse to fine.Specifically,based on Dynamic Time Warping and onset detection,a reference point and candidate point generation algorithm is designed to find a rough estimate of alignment.Secondly,an audio-score point pair screening algorithm based on the Support Vector Machine model is proposed to obtain accurate matching.Finally,a music theory based matching correction algorithm is designed to correct the alignment results.Experimental results on the BSAMS dataset demonstrate that the alignment accuracy increases by 32.5%on average compared with the traditional general audio-to-score alignment method.

关 键 词：音乐信息检索 音频乐谱比对 巴松演奏音乐 精准对齐 分段式 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]