An Emotional Symbolic Music Generation System based on LSTM Networks

With the development of AI technology in recent years, Neural Networks have been used in the task of algorithmic music composition and have achieved desirable results. Music is highly associated with human emotion, however, there are few attempts of intelligent music composition in the scene of expressing different emotions. In this work, Biaxial LSTM networks have been used to generate polyphonic music, and the thought of LookBack is also introduced into the architecture to improve the long-term structure. Above all, we design a novel system for emotional music generation with a manner of steerable parameters for 4 basic emotions divided by Russell’s 2-demonsion valence-arousal (VA) emotional space. The evaluation indices of generated music by this model is closer to real music, and via human listening test, it shows that the different affects expressed by the generated emotional samples can be distinguished correctly in majority.