I wonder if it's possible to fine-tune an image upscaling model on spectrograms, in order to clean up the sound?