Multi-class classifier of non-speech audio based on Fisher kernel