This really is just insanely cool. What a genius idea — take a machine-learning algorithm that can produce images from text, train it on images of spectrograms, let it interpolate between them, and convert the spectrograms back to audio. I could honestly listen to this for quite a while.