Long Short-Term Memory Recurrent Neural Network Architectures for Melody Generation

Recent work in deep learning has led to more powerful artificial neural network designs, including recurrent neural networks (RNNs) that can process input sequences of arbitrary length. We focus on a special kind of RNN known as a long short-term memory (LSTM) network. LSTM networks have enhanced memory capability which helps them in learning sequences like melodies. This paper focuses on generating melodies using LSTM networks and conducting a survey for verifying quality of melody generated. We used the Nottingham ABC Dataset, which is a database of over 14,000 folk songs in ABC notation and serves as the training input for our RNN model. We have also conducted a Turing Test to give the quality of the music generated by the model. We will also discuss the overall performance, design of the model and adjustments made in order to improve performance.