论文部分内容阅读
本文介绍了基于字符级长短期记忆网络(LSTM)计算机音乐生成方法,并分析了其不同网络结构在计算机音乐生成的效果.与现有的方法相比,基于字符级递归神经网络的音乐生成算法的优点是可以实现端到端训练,网络结构简单直观.本文方法的一种更为有意义的特征在于,音乐旋律无需通过繁琐的特征工程来获得,而是直接通过LSTM网络的学习自动获得.通过实验验证发现,不同LSTM网络结构在经过合适的训练之后均能够生成具有明显旋律的音乐序列.
This paper introduces computer music generation method based on LSTM and analyzes the effect of different network structures on computer music generation.Compared with the existing methods, the music generation algorithm based on character level recurrent neural network Has the advantage of end-to-end training and simple and intuitive network structure.An even more significant feature of this method is that music melodies need not be obtained through tedious feature engineering but are automatically obtained directly from the LSTM network. Experimental results show that different LSTM network structures can generate music sequences with obvious melody after proper training.