I am training up an audio recognition neural network using MFCCs. Its been shown to be very beneficial to artifically add noising and reverb of the original audio when training a neural network (Data augmentation). Up to now I've been noising the audio ahead of time, tripling, quadrupling or more the amount of space the source data takes up on disk.
It occurs to me however that it should be possible to add the noising to the MFCCs directly and save a vast amount of disk space (especially with lots of augmentations per file).
Most of the noising is simply additive. Is it possible to go from an MFCC to the filterbanks and then simply add the 2 together before converting back to MFCCs? Would this work? Would cepstral mean subtraction cause problems or would the mean subtraction need to be done after the combination?
If this would work then presumably convolutions (such as reverb) could also be done by simply multiplying the filterbanks together?
If it wouldn't work? Can you explain why?