基于WGAN和扩张卷积的符号音乐生成算法

Symbolic Music Generation Based on Dilated Convolution and WGAN

符号音乐生成模型大多采用循环神经网络结构递归生成,生成时随着时间推移顺序处理,存在缺乏空间相关性、并行性差、训练速度慢等问题。针对这些问题本文提出了一种新的音乐生成方法,即在2维空间域中基于Wasserstein距离生成对抗网络与深度卷积神经网络相结合的音乐生成算法,引入扩张卷积增加判别器的卷积核宽度来提取更多的旋律特征,并采用梯度惩罚实现Lipschitz约束,稳定网络训练。实验结果表明,模型有效地学习了音乐数据,生成的音乐符合人类作曲音符分布的自然性,且模型的收敛速度优于其他模型。

Most of the symbolic music generation models use Recurrent Neural Network(RNN)structure to generate,which is processed in sequence with the passage of time.There are some problems,such as lack of spatial correlation,poor parallelism,slow training speed and so on.To solve these problems,this paper proposes a novel music generation method,which is based on the combination of Wasserstein Generative Adversarial Network(WGAN)and deep convolution neural network in the music two-dimensional space domain.The method introduces the dilated convolution to increase the convolution core width of the discriminator to obtain more melody features,and uses gradient penalty to achieve Lipschitz limit,which stables GANs training.The experimental results show that the generator can learn the music data distributions effectively,the generated music accords with the naturalness of human composing note distribution,and the convergence speed of the model is better than other models.