Efficient Recurrent Neural Network Architecture for Musical Style Transfer

Mateusz Modrzejewski, Konrad Bereda, Przemyslaw Rokita

Abstract: In this paper we present an original method for style transfer between music tracks. We have used a recurrent model consisting of LSTM layers enclosed within an encoder-decoder architecture. In addition, a method for programmatic synthesis of sufficient, paired training datasets using MIDI data was presented. The representation of the data in the form of a real and an imaginary part of short-time Fourier transformation allowed for independent modeling of the music components. The proposed architecture allowed us to improve upon the state of the art solutions in terms of efficiency and range of applications while achieving high precision of the network.