Unsupervised learning of acoustic unit descriptors for audio content representation and classification