Super-Resolution for Music Signals Using Generative Adversarial Networks

Super-Resolution (SR) refers to increasing the resolution of a signal in a variety of ways, conventionally employed in the field of image enhancement. Compared with the endeavors for super-resolution in image processing, music signals require supper-resolution to improve their quality or adapt to communication in narrow-band channel, which is also regarded as bandwidth expansion. To this end, we shed light on super-resolution for music signals using the deep learning strategy of Generative Adversarial Networks (GANs). The proposed approach feeds Shot-Time Fourier Transform (STFT) features of low-band signals to the GAN, expecting to obtain their high-band information through jointly considering content and adversarial losses. Then, we carry out experiments on MUSDB18 dataset using mixtures of music sources, in order to show the performance of the proposed approach. The experimental results indicate that the proposed approach achieves better super-resolution performances compared with interpolation and some conventional deep-learning strategies.