We show how we can fit acoustic models of objects from real world audio of impact sounds, even from audio taken from ASMR YouTube videos. Robots can use our models to separate sound sources of impacts between multiple objects, then identify distinct properties of each object.