Deep Learning Approaches for Melody Generation: An Evaluation Using LSTM, BILSTM and GRU Models

Music generation is an application of machine learning that has garnered significant attention over the recent past. In this study we generated musical notes using three deep learning models- (LSTM) Long Short-Term Memory, (BiLSTM) Bidirectional LSTM and (GRU) Gated Recurrent Unit. We used the classical piano dataset which consists of 295 MIDI files with a diverse range of piano pieces to train our models. These are represented as sequence of events such as MIDI notes which are fed into the models. The model generates a new sequence of notes based on the pattern it has learned from training data. We found that LSTM outperforms the other two models in terms of accuracy and musical quality with a training accuracy of 85.77% and validation accuracy of 77.89%. A training accuracy of 77.89% and a validation accuracy of 65.52% are shown by the BILSTM model, in contrast. The GRU model displays training and validation accuracy as 81.36% and 66.54%, respectively. Since LSTM produces sequences most effectively, using it would be the best.