Generate Polyphonic Music with Multivariate Masked Autoregressive Flow

This paper explores the use of the Masked Autoregressive Flow (MAF) model for music generation, specifically addressing its limitation to univariate time series.To extend MAF for polyphonic melodies, three approaches are proposed and tested on the Lakh Pianoroll Dataset.The results show promising accuracy and the model's ability to generate original, pleasing melodies, demonstrating the potential of this innovative interdisciplinary approach.