基于深度学习的音乐源分离算法研究综述  

A Review of Music Source Separation Algorithms Based on Deep Learning

作　　者：林东星 桂文明[1] 胡夏芸 李相龙 张静兰 LING Dongxing;GUI Wenming;HU Xiayun;LI Xianglong;ZHANG Jinglan(Jinling Institute of Technology,Nanjing 211169,China;Queensland University of Technology,Brisbane 4072,Australia)

机构地区：[1]金陵科技学院软件工程学院,江苏南京211169 [2]昆士兰科技大学,昆士兰州布里斯班4072

出　　处：《金陵科技学院学报》2025年第1期22-32,共11页Journal of Jinling Institute of Technology

基　　金：教育部“春晖计划”国际合作科研项目(HZKY20220134);江苏省现代教育技术研究智慧校园专项课题(2023-R-107296);教育部产学合作协同育人项目(230805181044709);江苏省高等学校自然科学研究重大项目(22KJA520001)。

摘　　要：详细介绍了基于深度学习的音乐源分离方法,并分别阐述频域、时域和混合域三种音乐源分离模型的原理、优势和成果。首先,叙述了客观评价方法的重要性,描述了源失真比、源干扰比、源伪影比等客观指标,用于量化分离效果。然后,阐述并比较了几个重要的数据集。最后,分析了该领域目前面临的挑战,主要包括信号时频域混叠导致分离难、深度学习模型计算复杂、现有指标难以评价音乐的主观感受。针对上述问题,本文给出了相应的解决策略,并给出了未来的研究方向,为音乐源分离领域的进一步研究提供了参考。The paper introduces in detail the music source separation methods based on deep learning,and presents the principles,advantages,and achievements of three types of methods:frequency domain,time domain,and hybrid domain approaches.First,it describes the importance of objective evaluation metrics and introduces objective indicators such as source distortion ratio(SDR),source interference ratio(SIR),and source artifact ratio(SAR)to quantitatively measure the separation effect.Subsequently,several important datasets are discussed and compared.Finally,some current challenges in this field are analyzed,including the difficulties of separation caused by the aliasing of signals in the time-frequency domain,the high computational complexity of deep learning models,and the difficulty of using existing indicators to evaluate the subjective perception of music.To address these issues,this paper proposes corresponding solutions and outlines future research directions,providing a reference for further research in field of music source separation.

关 键 词：深度学习 音乐源分离 评估指标 数据集 

分 类 号：TN912.1[电子电信—通信与信息系统]