Allow calibration to happen after pre-processing and (maybe) feature extraction.

[damellis]

This would make it easier, for instance, to do audio calibration in the feature domain (after FFT and FFT feature extraction) rather than the time domain.