Do sound event representations generalize to other audio tasks? A case study in audio transfer learning