i vaguely remember projects about preprocessing your file collection using ML (on the audio content) to build your own similarity index to run local recommendations