基于生成对抗网络的智能音乐制作综述

如何借助计算机算法进行音乐的自动或半自动化生成工作一直是人工智能领域的一个研究热点。近年来,随着深度学习技术的深入发展,使用基于神经网络并契合乐理先验知识的方法来生成高质量、多样性智能音乐的任务也引起了研究者的重视。其中,引入生成对抗机制以提升生成效果的工作取得了一定成果,同时也具备极大的提升空间。为了更好地推进后续研究工作,对相关领域的现有成果进行全面而系统的梳理、分析、总结具有比较重要的意义。首先对机器作曲的发展过程进行了回顾,对音乐领域常用的GAN相关重要模型进行了简要归纳介绍,对引入了生成对抗训练机制的音乐生成方法进行了重点分析,最后对该领域的现状进行了总结,并进一步展望了未来的发展方向。

Survey of intelligent music creation based on GAN

Recently,researchers pay more and more attentions for automatic or semi-automatic music generation based on computer algorithm.With the further development of deep learning,researchers start to focus on producing high-quality and multifarious-style music with neural network and priori knowledge of music theory.Furthermore,several works introduced generative adversarial network(GAN)to try to improve the quality of the results.To summarize the important results in this area is meaningful and with guiding significance for the following works.This paper firstly reviewed the history of intelligent music,then listed related GAN model that are commonly applied in music creation,after that it analyzed some important works in this area.Finally,it presented some observations,and prospected future work direction.