Generation of Western Piano Music Based on Deep Learning

Abstract

A large number of automatic composition models based on deep learning have been proposed in the field of artificial intelligence. This paper regards music as a series of sequences, and proposes an improved structure of transformer (RM- Transformer), which uses random mask module to replace the original mask module. Firstly, music features are extracted during data preprocessing, and then the processed data is input to RM Transformer for training. This model learns the music features contained in the data itself. Finally, music can be generated using the trained model and compared with other network models. Among them, the prediction accuracy and sequence similarity increased by 6.6% and 9.6% respectively, and the harmony and melody of music have been greatly improved. The network structure is more suitable for music generation.