Service for Music Synthesis

Within this paper, there were studied and implemented methods for solving the problem of transferring the style of music composition. This task is not as well known as the generation of music by a neural network, but it becomes more and more relevant: its solutions were presented in 2018 and 2019. For data preprocessing there were considered the following approaches: analysis of all notes in a composition, saving only notes that make up the main melody tune and tracks with percussion instruments. There were tested the following methods of dataset creation: matching each note with the next in composition and finding the most similar note in the already processed version. As models of machine learning, there were taken several machine architectures with such recurrent layers as LSTM and GRU, as well as with a one-dimensional convolutional layer Conv1D. In order to provide access to the developed solution, there was created a server with a simple API that allows melodies to be processed in some of the genres presented. During the experiments, it was revealed the dependence of the melody synthesis results on the selected data preprocessing methods and dataset creation. Keywords: machine learning, music processing, style transfer.