基于深层神经网络的音乐自动生成  

Automatic music generation based on deep neural network

作　　者：周林[1] ZHOU Lin(Preschool Normal Department University for Science&Technology,Meishan 620564,Sichuan Province,China)

机构地区：[1]四川科技职业学院幼儿师范学院,四川眉山620564

出　　处：《信息技术》2022年第5期53-61,共9页Information Technology

基　　金：国家自然科学基金(61972268)。

摘　　要：人工智能尽管在图像生成和短序列生成领域有着非常卓越的表现,但对于符号音乐的生成仍然有许多不足之处。因此,通过基于LSTM和WaveNet两种模型解决旋律生成的受限问题,探讨音乐结构的显式结构编码效果。实验结果表明,使用一堆扩展的卷积层对结构进行更明确的编码可以显著地提高性能,并且在生成过程中对底层和弦级数进行全局编码可以获得更好的效果。Although artificial intelligence has excellent performance in the field of image generation and short sequence generation,it still has many shortcomings for the generation of symbolic music.By improving the two models based on LSTM and WaveNet,solve the problem of limited melody generation,the effect of explicit structure coding of music structure is discussed.Experiment results show that using a bunch of expanded convolutional layers to encode the structure more clearly can significantly improve performance,and global encoding of the underlying chord progression can achieve better results during the generation process.

关 键 词：音乐生成 人工智能 机器学习 音乐结构分析 全局编码 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]