Unsupervised learning to extract high level features might work. Then get examples of the kinds of sounds people are training it on and use that to extract even better features.