To do that you'll need a way of generating similar whispers.  This
   means simulating cover noise overlayed on the DeepSpeech training data.

   Obviously this idea has the error that the training data is not
   whispers.  So you'll also want to train something to classify data as
   whispers or not.