Improve audio featurizer by adding e.g. resampling, db_normalization, and random shift, as suggested in speech_dl codes.