Towards Music Generation With Deep Learning Algorithms

Abstract

Computer music generation has application in many areas, including computer aided music composition, on demand music generation for video games, sport events, multi-media experiences, creating music in the style of passed away artists, etc. In this work we describe our approach towards music generation. We trained a deep learning model on a corpus of works of several authors. By priming the model with a snippet of an authors work we used it to create new music in their style. The dataset consists of music for guitar in midi format, containing only 1 part/instrument. We gathered more than 2000 files, of which we used from 5 to 300 per experiment. The data for the deep learning model is represented in piano roll format, a binary matrix where one axis represents the time and the other axis represents midi notes. Two deep learning architectures were evaluated, a 2-layer recurrent neural network of LSTM (Long Short Term Memory) cells and an Encoder-Decoder (Auto-Encoder) architecture for sequence learning, where both the encoder and decoder are built as recurrent layers of LSTM cells. The models were implemented in the Keras deep-learning library. The results were evaluated on a subjective basis, and with the evaluated datasets both architectures produced results of limited quality.