Generating Musical Compositions with GRU and LSTM Neural Networks

Abstract

The field of music generation has witnessed remarkable advancements in recent years, thanks to the emergence of deep learning techniques. In this paper, we present a novel music generation system utilizing a character level Recurrent Neural Network (char RNN) empowered by a hybrid architecture of Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) cells. Our system is trained on a comprehensive dataset consisting of 340 tunes represented in ABC notation, a text based format for musical compositions. By predicting the subsequent character in a sequence of ABC notation, our model adeptly generates new and captivating melodies. Extensive evaluations are conducted, employing a variety of metrics, to assess the system’s performance, revealing its ability to generate coherent and musically plausible music. Furthermore, we demonstrate the versatility of our proposed system, highlighting its potential applications in music composition, accompaniment, and real time improvisation. The results substantiate the effectiveness of employing char RNN with LSTM GRU cells for music generation, thereby opening up intriguing avenues for future research in this evolving field.