Automatic Music Generator Using Recurrent Neural Network

Abstract

In this paper, we developed an automatic music generator with midi as the input file. This study uses long short-term memory (LSTM) and gated recurrent units (GRUs) network to build the generator and evaluator model. First, a midi file is converted into a midi matrix in midi encoding process. Then, each midi is trained on a single layer and double stacked layer model of each network as a generator model. Next, classification model, based on LSTM and GRU, are trained and chosen as an objective evaluator to analyze the performance of each generator model which classify each midi based on its musical era. Subjective evaluation is conducted by an interview with volunteer respondents with various backgrounds such as classical music interest, performance, composer, and digital composer. The result shows that the double stacked layer GRU model perform better to resemble the composer pattern in music with 70% score of recall. Moreover, subjective evaluation shows that the generated music is listenable and interesting with the highest score of 6.85 out of 10 on double stacked layer GRU.