Music Generation using Deep Generative Modelling

摘要

Efficient synthesis of musical sequences is a challenging task from a machine learning perspective, as human perception is aware of the global context to shorter sequences as well of audio waveforms on a smaller scale. Autoregressive models such as WaveNet use iterative subsampling to generate short sequences that are a result of a localized modeling process but lacking in overall global structures. In juxtaposition, Generative Adversarial Networks (GANs) are effective for modeling globally coherent sequence structures, but struggle to generate localized sequences. Through this project, we aim to propose a system that combines the random subsampling approach of GANs with a recurrent autoregressive model. Such a model will help to model coherent musical structures effectively on both, global and local levels.