Robust speech recovery of lossily-encoded audio and speech by integrating and embedding domain-specific signal models within deep-learning frameworks