A transformer generative adversarial network for multi‐track music generation

Abstract

Abstract This study proposes a new generation network based on transformers and guided by the music theory to produce high‐quality music work. In this study, the decoding block of the transformer is used to learn the internal information of single‐track music, and cross‐track transformers are used to learn the information amongst the tracks of different musical instruments. A reward network based on the music theory is proposed, which optimizes the global and local loss objective functions while training and discriminating the network so that the reward network can provide a reliable adjustment method for the generation of the network. The method of combining the reward network and cross entropy loss is used to guide the training of the generator and produce high‐quality music work. Compared with other multi‐track music generation models, the experimental results verify the validity of the model.

Keywords

• artificial intelligence
• deep learning
• multimedia