基于变分自编码器的交互式旋律生成方法

现有的交互式神经音乐生成方法主要存在控制模式不灵活、数据标注困难以及模型难以优化等问题。针对这些问题,提出了一种基于变分自编码器(VAE)的无监督交互式旋律生成方法。通过为VAE引入显式的旋律轮廓条件推理学习,实现了对生成旋律局部与全局特征的灵活控制。实验表明,该方法易于优化且具有良好的旋律局部与全局特征的控制能力。通过对大量生成样本的分析,证明了模型从音乐数据中学习到了有用的音乐知识。

Interactive melody generation based on variational autoencoder

The existing interactive neural music generation methods were inflexible in control modes,high cost in manual labeling and difficult to optimize.By analyzing these methods,this paper proposed an unsupervised interactive melody generation method based on VAE.By introducing explicit conditional inference learning of melody contour to VAE,this method realized the flexible control to local and global features of generated melodies.Experiments show that this method is easy to optimize and obtain good control ability for local and global melody features.By analyzing generated samples,it proves that the model can learn useful music knowledge from music data.