All for suggestions! We've tried white noise as well, but it only works on plain talking samples (not music, for example). My guess is that the most robust solution will come from updating how it's trained.