LSTM based Music Generation with Dataset Preprocessing and Reconstruction Techniques

摘要

Numerous approaches have been used by researchers for the purpose of music generation. Recurrent neural networks (RNNs) and Long short term memory (LSTM) networks are able to effectively model sequential data. LSTM networks have been extensively used to produce sheet music, character by character. These LSTM models, however, require a lot of time to train to be able to produce pleasant and syntactically correct sheet music. We introduce some effective dataset preprocessing and reconstruction techniques which facilitate the generation of syntactically correct sheet music, while reducing the training time. The quality of music generated is qualitatively measured by peers. The proposed model employing the dataset preprocessing and reconstruction techniques is compared with another model possessing no such techniques in a subjective manner.