Towards an Evaluation of Symbolic Music Encodings for RNN Music Generation

The choice of encoding for symbolic music data used in music generation models has typically been done in a case-by-case basis up to now. In this paper we attempt to evaluate and study the behaviour of a baseline shallow LSTM network trained on Irish folk music, using three different encodings: a MIDI-like event-based encoding, a fixed timestep-based encoding, and the popular ABC notation. We use objective statistical measures on the network output and also visualise the LSTM parameters to gain insight on the way it processes each encoding.