This is just one such site, using the stable diffusion model at the core but using it to draw the spectrogram of prompted music, then using some basic musical constraints to keep a steady tempo, and inverse short-time fourier transforms to convert it to time-domain audio.