Improving Automatic Jazz Melody Generation by Transfer Learning Techniques
Paper: Improving Automatic Jazz Melody Generation by Transfer Learning Techniques
Authors: Hsiao-Tzu Hung, Chung-Yang Wang, Yi-Hsuan Yang and Hsin-Min Wang
Comments: Accepted to APSIPA 2019.
Abstract: In this paper, we tackle the problem of transferlearning for Jazz automatic generation. Jazz is one of rep-resentative types of music, but the lack of Jazz data in theMIDI format hinders the construction of a generative modelfor Jazz. Transfer learning is an approach aiming to solve theproblem of data insufficiency, so as to transfer the commonfeature from one domain to another. In view of its success inother machine learning problems, we investigate whether, andhow much, it can help improve automatic music generation forunder-resourced musical genres. Specifically, we use a recurrentvariational autoencoder as the generative model, and use agenre-unspecified dataset as the source dataset and a Jazz-onlydataset as the target dataset. Two transfer learning methods areevaluated using six levels of source-to-target data ratios. The firstmethod is to train the model on the source dataset, and thenfine-tune the resulting model parameters on the target dataset.The second method is to train the model on both the sourceand target datasets at the same time, but add genre labels tothe latent vectors and use a genre classifier to improve Jazzgeneration. Our subjective evaluation shows that both methodsoutperform the baseline method that uses Jazz data only fortraining by a large margin. Among the two methods, the firstmethod seems to perform better. Our evaluation also shows thelimits of existing objective metrics in evaluating the performanceof music generation models.
