基于深度学习的自动乐谱生成系统  

Deep learning based automatic sheet music generation system

作　　者：刘丹霞 路惠捷 LIU Danxia;LU Huijie(College of Basic Medical Sciences,Air Force Medical University,Xi’an 710032,China;Department of Military Psychology,Air Force Medical University,Xi’an 710032,China)

机构地区：[1]空军军医大学基础医学院,陕西西安710032 [2]空军军医大学军事医学心理学系,陕西西安710032

出　　处：《电子设计工程》2025年第6期21-24,29,共5页Electronic Design Engineering

基　　金：陕西省科技厅一般项目(2022GY-072)。

摘　　要：为解决乐谱资源获取困难的问题,研发了一种基于深度学习的自动乐谱生成系统。该系统利用门控循环单元和长短期记忆网络,分别构建了音源分离模块和和弦估计模块。音源分离模块通过增加可分离音源数量提高了效率,而和弦估计模块则能准确识别和弦。在DSD100数据集上的音源分离模型训练结果显示信号分离比率显著提升,自建数据集上的和弦估计模块的准确率显著提高。该系统为音乐创作者和爱好者提供了高效的乐谱自动生成工具。In order to solve the problem of difficulty in acquiring musical score resources,an automatic musical score generation system based on deep learning was developed.The system utilises gated loop units and Long Short-Term Memory networks to construct sound source separation and chord estimation modules,respectively.The sound source separation module improves efficiency by increasing the number of separable sound sources,while the chord estimation module accurately recognises chords.The training results of the source separation model on the DSD100 dataset show a significant improvement in the signal separation ratio,and the chord estimation model on the self-constructed dataset achieves a significant improvement.The system provides music creators and enthusiasts with an efficient tool for automatic score generation.

关 键 词：深度学习 RNN LSTM 乐谱 音源分离 和弦估计 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP391[自动化与计算机技术—控制科学与工程]