Folk melody generation based on CNN-BiGRU and Self-Attention

Using various deep learning networks to generate music is a research hotspot in the field of human intelligence. Due to the limitation of network structure in the existing melody generation models, the quality of melody generation is poor. In this paper, for folk melody, we propose a melody generation network based on CNN-BiGRU and Self-Attention. First, the processed data is input into the CNN module consisting of one-dimensional convolution layer and pooling layer to extract local features. Then, the features extracted by CNN module are constructed as time series and is input into BiGRU network to extract global features. Finally, Self-Attention assigns different weights to the features of BiGRU output to highlight important feature information during note generation. Experimental results show that the prediction accuracy of the proposed model is improved by 3.61 % compared with BiGUR, 5.26% compared with CNN, and 6.53% compared with CNN-BiGRU. Meanwhile, the model presented in this paper also achieves improvement in other evaluation measures.

I. Introduction

Automatic melody generation means that melody can be automatically generated quickly by learning the theoretical rules of composition through computer[I]. Automatic melody generation can assist the composer to create music. At the same time, it can help music lovers compose their own melodies and lower the threshold of melody creation[2].