基于变分自编码器的音乐旋律生成算法
首发时间:2020-12-23
摘要:当前用于音乐旋律生成的深度生成算法大部分基于对抗神经网络(Generative Adversarial Networks,GAN)或变分自编码器(Variational Autoencoder,VAE)。本文着重研究了基于VAE的音乐旋律生成算法,旨在改善VAE的优化目标以提高生成质量。在SeqGAN的启发下,本文提出了VAE的一个变种,称其为SeqVAE,它为VAE引入了强化学习中的策略梯度。引入策略梯度的SeqVAE可被认为是一种VAE和GAN的结合,相较于普通的VAE因添加了策略梯度损失而减少了优化目标的偏差,从而在一定程度上提高了学习能力。最终由实验可得,SeqVAE要优于基线,能够生成旋律性更强、节奏变化更优美的旋律。
关键词: 计算机应用技术 深度学习 变分自编码器 音乐旋律生成
For information in English, please click here
Music melody generation algorithm based on variational autoencoder
Abstract:Most of the current depth generation algorithms used for music melody generation are based on Generative Adversarial Networks (GAN) or Variational Autoencoder (VAE). This paper focuses on the research of music melody generation algorithm based on VAE, aiming to modify the optimization goal of VAE to improve the quality of generation. Inspired by SeqGAN, this paper proposes a variant of VAE, called SeqVAE, which introduces a policy gradient in reinforcement learning. SeqVAE can be considered as a combination of VAE and GAN. Compared with plan VAE, it reduces the deviation of the optimization goal due to the addition of policy gradient loss, thereby improving the learning ability to a certain extent. In the end, it can be obtained from experiments that SeqVAE is better than the baseline and can generate a more melodic melody.
Keywords: technology of computer application deep learning variational autoencoder music melody generation
基金:
引用
No.****
动态公开评议
共计0人参与
勘误表
基于变分自编码器的音乐旋律生成算法
评论
全部评论0/1000