Robust localisation of multiple speakers exploiting head movements and multi-conditional training of binaural cues.