Text this: Relevance-based quantization of scattering features for unsupervised mining of environmental audio.