Automatic music composition could dramatically decrease music production costs, lower the threshold for the non-professionals to compose as well as improve the efficiency of music creation. In this paper, we proposed an intelligent music composition neutral network to automatically generate a specific style of music. The advantage of our model is the innovative structure: we obtained the music sequence through an actor’s long short term memory, then fixed the probability of sequence by a reward-based procedure which serves as feedback to improve the performance of music composition. The music theoretical rule is introduced to constrain the style of generated music. We also utilized a subjective validation in experiment to guarantee the superiority of our model compared with state-of-the-art works.

自动音乐创作可以大大降低音乐制作成本，降低非专业人士创作的门槛，并提高音乐创作的效率。在本文中，我们提出了一种智能音乐创作中性网络，可以自动生成特定风格的音乐。我们的模型的优点是创新的结构：我们通过演员的长期短期记忆获得了音乐序列，然后通过基于奖励的过程固定了序列的可能性，该过程可作为反馈来改善音乐创作的性能。引入了音乐理论规则来约束所生成音乐的风格。我们还在实验中采用了主观验证，以确保我们的模型与最新技术相比具有优越性。