Humming-to-Instrument Conversion based on CycleGAN

Abstract

In this research, we propose a humming-to-instrument conversion system based on cycle-consistent adversarial networks (CycleGAN) to convert human humming to the sound of viola. This research adjusts the weight of cycle-consistency loss and identity loss to successfully convert the sound. From the experimental results of objective RMSE, the converted audio is more similar to viola compared to the similarity to humming. From the results of subjective MOS, the quality of the converted sound is fair to listeners.